Subscribe now

My first day with the bioreactor, I was going to infect HEK293 cells with a stock of adenovirus, to turn these thriving, living cells in the bioreactor, into factories which assemble more viruses. These would eventually be concentrated to turn into a vaccine. I was hooked onto this process of “optimizing” the growth of cells in this small glass world to somehow create a perfect Goldilocks environment for the cells.

Bioreactors are widely used in the pharmaceutical industry for various applications. Viral vaccines, gene therapies, cell therapies, recombinant proteins, and other biologics. It’s an optimized, engineered world where cells can grow by maintaining pH, dissolved oxygen and nutrients. The critical aspect of the bioreactor design is scaling it from the bench-top size to large scale industrial sizes (eg. 2L to 50L) while maintaining the same biological environment. When a process is being optimized at bench scale, parameters which cells care about have been tweaked to derive the best culture performance in terms of therapeutic product and cell density. However a 2L and a 50-L bioreactor do not behave identically even when you keep the same pH, temperature and oxygen concentration and most of the engineering in this space is to ensure that the scale up can happen in the best possible way.

This is why there are a lot of startups that are building novel technologies in this space to help close the gaps.

1. Sunflower Therapeutics

Sunflower Therapeutics has developed the first microbial perfusion fermentation system designed and built for continuous protein production, known as Daisy Petal. Their core objectives includes making continuous production hassle-free, and to reduce the footprint size of biomanufacturing platform to make the production of therapeutics more cost-friendly. The platform integrates continuous fermentation, with chromatography for seamless bioprocessing capability. Sunflower Therapeutics is a women-owned public benefit corporation and has raised a total of $5.64M over 5 funding rounds: 1 Seed, 2 Early-Stage and 2 Grant (prize money) rounds.

2. Bioscibex

   Bioscibex, is a Swiss company that is developing a single use bioreactor that can simplify the 7 step cell culture process to 2 steps, thereby reducing the potential for contamination, and making the process safer, simpler and less manual. It’s flagship platform, SW1NGO, is designed to simplify the traditional cell culture "seed train" by enabling expansion from 30 mL to 30 L within a single disposable bioreactor bag. Bioscibex has raised a total of CHF 1 million, supplemented by CHF 150 000 from Venture Kick.

   

3. BioThrust

   
   BioThrust, a German startup has developed the ComfyCell, which is a low-shear membrane-stirrer bioreactors for iPSCs, MSCs, NK/T cells, organoids and viral vectors. It’s proprietary membrane stirrer replaces conventional bubble sparging with diffusive, bubble-free gas transfer through hollow-fiber membranes. According to the company, this approach delivers high oxygen transfer while substantially reducing shear stress and eliminating foam. Their core objective is to enable scaling of sensitive cells without damaging them. BioThrust has raised a pre-seed extension funding round on April 24, 2026, following earlier investments like a €2 million grant via GO-Bio next in December 2025 and a seven-figure investment from Freigeist Capital in January 2024.

4. Lyric Bio

Lyric Bio is developing a next-generation biomanufacturing platform that allows human tissue growth in large 3D cell culture that replicates the cell density of human tissue. The company is looking into human Immunoglobulin (IVIg and SCIg). Lyric Bio has created quite a buzz in the industry, with its oversubscribed seed round which was led by the The Venture Collective and with participation from Black Diamond Ventures, Draper Associates, SOSV, Lucas Venture Group.

Bioreactor scale up is still a major challenge in the industry. It requires extensive experimentation, expertise, and a sound strategy based on science. It’s a problem that still has a lot of room for innovation and better engineering, especially to reduce the complexity of scale-up. The more I think about it, bioreactor scale-up really feels like a game of musical chairs with physics. When you try to keep one parameter constant, another one gets knocked out of place. The objective then becomes the identification of chairs which matter the most to keep the miniature world of cells from crashing down.

Hi there,

Welcome back to another edition of Startup Intel! I am experimenting with new ways to make this very useful and trying a new format for these briefings. You may also have noticed that I am covering broader sectors now - biotech / digital health / healthcare. Nevertheless, the goal is still to bring the critical updates in these sectors to your inbox. Hope you enjoy reading this week’s post!

Cheers,

Divya

»What’s the Buzz?

FDA Approval of the First Standalone Robotic Blood Draw Device

The FDA has approved the first standalone robotic blood-draw device- Aletta in the U.S., allowing the device system to autonomously perform the entire venipuncture process as long as there is oversight from a phlebotomist. Aletta was developed by Vitestro, which is a Netherlands-based medical robotics Series B startup. It has raised $111 million to date from prominent investors, including the Mayo Clinic, Labcorp Venture Fund, and Sonder Capital.

Vitestro hopes to make the patient experience of blood draw, smoother and reduce wait times caused by phlebotomist shortages.

The Aletta can handle every step the blood draw, while a trained phlebotomist will have to begin each session and remains available to respond to any issues throughout. The device guides the patient to position their arm, after which the patient or supervisor presses a button to start the blood draw. The Aletta then uses near-infrared light and Doppler ultrasound to locate a suitable vein and tell it apart from arteries. If no appropriate vein is found, the device will not attempt the procedure.

Once a vein is identified, the Aletta proceeds through the remaining steps on its own: applying a tourniquet, preparing the skin, inserting and disposing of the needle, changing collection tubes and placing a bandage.

Pic Credit: Vitestro

Revolution Medicine’s Big Pancreatic Cancer Win

On August 26, the FDA approved daraxonrasib (Rasonque). The groundbreaking RAS GTPase inhibitor doubled survival rates in patients with advanced metastatic pancreatic adenocarcinoma during its clinical trials.

This was noted as a monumental moment for the treatment of pancreatic cancer and Rasonque is expected to become a blockbuster drug very soon. RAS-addicted cancers are malignancies that depend entirely on mutated RAS genes. For a long time, researchers considered KRAS proteins, the primary genetic driver behind tumor growth to be entirely "undruggable". Revolution Medicine has successfully developed a line of RAS(ON) inhibitors capable of locking onto these proteins and turning off the tumor’s growth signals.

New Partnerships

Lady Gaga and Michael Polansky officially launched Outer Biosciences. The company raised $23 million to combine AI modeling with living human skin tissues to discover new compounds for regenerative medicine and cosmetics. The company states that their biorxiv paper : https://www.biorxiv.org/content/10.64898/2026.04.30.721933v1 serves as the foundation for their work. Their platform technology includes the full stack - machine learning, novel target discovery, robotics.

Overall, it’s been a very happening week for the biotech world!

Funding Deadlines

• FastForward Accelerator deadline is coming up fast : 7 Sep 2026 https://www.ffwd.org/accelerator

• Entrepreneur First deadline for San Francisco Bridge Residency Fall 2026 is 30 August 2026. Cohort starts in October 26.

As I was scouting for startups in the longevity space, I happened to come across HexemBio on Linkedin. I messaged the founder of HexemBio, Gabriel, not really expecting a response. Nevertheless, he got back to me, met me for a good 30 minutes, patiently explaining what the technology they are building is all about. It was impressive and his enthusiasm was palpable. HexemBio stood out to me as a company working at the very edge of what science can do to solve one of humanity’s fundamental challenges - getting old.

HexemBio has raised $10.4M in seed funding led by Draper Associates, along with SOSV, Seraphim and other investors. It has a stellar team made up of researchers from the most elite universities and research programs. HexemBio's advisory board includes Robert S. Langer, PhD, Institute Professor at MIT and Co-Founder of Moderna. Their core technology is based on the paper they have published in Nature : Modelling post-implantation human development to yolk sac blood emergence. For this technology, HexemBio utilizes human pluripotent stem cells. These cells have the capability to develop into any kind of cells in the body. Using these special cells, they are able to guide them genetically to form embroyo models invitro, and to even develop a functioning yolk sac, which is responsible for creating the very first blood cells. Creating this “synthetic primitive” state, has an advantage. Placing adult aged stem cells in this primitive environment biologically resets them to environment t restoring their native regenerative power, as if they magically remember how to. be young again.

HexemBio received FDA’s Orphan Drug Designation in July 2025 for their lead program, for improving outcomes for bone marrow transplants. They also have a broader vision to combat immune aging, neurodegeneration, and disease to extend human healthspan.

HexemBio’s tagline on their website says - Live longer, healthier - happier. The ultimate dream that every human strives to achieve. While it seems like a utopian fantasy at first glance, HexemBio’s impressive technology has made me pause and ask, Why not? If this technology improves healthspan, I am more than happy to keep my rose-tinted glasses on.

For years, quantum computing seemed like a promised future that was mostly confined to academic labs. A science-fiction that was far away from being useful for real world applications. IonQ is changing that today to build something that is practical for real-world applications. I spent sometime diving into what this company is up to, and here’s why I think it’s really worth paying attention to them.

Subscribe now

IonQ has made several major acquisitions recently :

• SkyWater Technology: Acquired semiconductor foundry SkyWater Technology to bring chip fabrication and trap manufacturing entirely in-house. This makes it the only full-stack quantum platform company. They are also the largest, exclusively US-based semiconductor foundry. This significantly reduces the risk of geopolitical issues or supply chain vulnerabilities affecting availability of chips.

• Skyloom Global Acquisition : To expand quantum-secure optical networking.

• Oxford Ionics : Integrates Oxford Ionics' Electronic Qubit Control (EQC) technology, which controls trapped ions using standard semiconductor chip circuitry rather than external lasers.

• Vector Atomics: Expands its portfolio into specialized portfolio in advanced quantum sensing, precision timing, and positioning hardware.

• ID Quantique : ID Quantique, is a leader in quantum cybersecurity and this acquisition adds 300 new patents to IonQ’s portfolio.

• Seed Innovations : To expand its software and AI capabilities.

IonQ also had several technical breakthroughs recently that gives a glimpse of where quantum computing is headed. Quantum computers are limited in their capability today for computation mainly because of their error rates. We are in the NISQ (Noisy-Intermediate Scale Quantum) era. The error rates are high and for precise calculations, where errors are detrimental, this limits their capability to be used in large scale systems. One of the work arounds that IonQ and other quantum companies use are hybrid algorithms which use both classical and quantum algorithms in a loop. Recently, IonQ published a paper on arxiv : The Walking Cat Architecture. This design blueprint for the next generation of quantum computers was path-breaking. It was the outline for how trapped-ion quantum computers can overcome the biggest pain point that quantum computers have always faced : Errors. IonQ is progressing at a faster rate towards fault-tolerant quantum computers. Technology that was initially thought to be 20 years away, seems to be closer. IonQ is with industry partners from various fields to show how QC can make an impact in their research workflows. As for biotech, IONQ has already partnered with Astra Zeneca, to demonstrate the largest quantum-accelerated electronic structure simulation executed on quantum hardware, specifically modeling the activation barriers of a Suzuki-Miyaura reaction.

Why am I bullish on quantum computers? Classical computers will reach a bottleneck when computing more complex systems. Approximations and course-grained models will not be sufficient to design complex systems. As we progress towards solving the harder challenges in biotech or chemistry, fault-tolerant quantum computers will hold the key to developing better drugs, designing better reactions, and optimizing processes in ways in which classical computers cannot keep up. For now, I am watching IonQ closely to see how it changes biotech drug development.

Share

Share Divy's Dispatch

Leave a comment

Subscribe now

Hey there,

I’m so glad you’re back here again! Here is this week’s Startup Intel to update you on everything exciting, science-fictiony, and disruptive happening in the startup world. Think of this as a weekly startup newspaper (yes i’m old school) :) Also, if you are a startup founder, I’d love to hear from you. Hit the button below and tell me what you are working on. I want to feature early-stage builders in this newsletter. For everyone else, if there is anything you wish this newsletter featured, please let me know. I’d love to hear your feedback to make this newsletter as informative and relevant for you as possible!

Cheers,

Divya

» What’s the Buzz

• Skylark Bio , a clinical stage biotech startup emerges from stealth on 11 Aug, announcing that they have dosed their first patient with a therapy aimed at restoring hearing in a child with a GJB2 mutation. This is a closely watched space as there are no existing treatment options for deafness caused by this mutation.

• Astra Zeneca - Bristol Myers Squibb Merger Talks : The Financial Times reported the two were involved in early talks to create a $400B pharma giant, but rumors were quickly dismissed by both companies. Almost £17B was wiped off the value of AstraZeneca as soon as the news of the merger talks emerged.

» New Deals

• Boulevard Bio emerges from stealth with $65 million from founding investor Deerfield Management to advance a broad pipeline of multi-specific antibodies for B cell-driven autoimmune diseases.

• Aureka Biotechnologies has announced the closing of US$100 Million Series B funding to build a biological world model for drug discovery. It is developing a Lab-in-the-Loop system at the core, which is used for model development, incorporating AI agents, high-throughput digital biology, proprietary single-cell functional screening and its proprietary in-house experimental platform.

» Get Funded

• 500 Global (Founders Fund Track): September 1, 2026 (Late-seed to early Series A).

• Techstars (Global Fall Cohorts): August 15, 2026. The deadline for standard entry across its 40+ specialized vertical tracks.

• On Deck (Pre-Seed Track): August 30, 2026. (Pre-product, first time founders)

» Trends and Market Analysis

A key trend that I am noticing is that tech giants are moving from just software into hardware. Companies are looking into generating their own biological datasets and using it to improve their models, and incoporate automation and high throughput wet lab capabilities. It is now quite evident that the publicly available datasets have reached a saturation, and true moat for any company lies on how much it can fully integrate into the biological discovery workflow.

» Now Hiring

Superluminal Medicines is now hiring Director / Senior Director, CMC Drug Substance (Boston, MA| Posted on 14 Aug)

Maze Therapeutics is now hiring a Senior Project Manager, Clinical Operations (South San Francisco, CA| 14 Aug)

Terray Therapeutics is hiring an Associate Director of Product Management (Remote (United States)| 13 Aug)

Systimmune is hiring Senior Clinical Research Associate - Oncology (United States |13 Aug)

Speaking of breakthroughs, I can't imagine a more exciting startup to write about than Varda Space Industries. In today’s Startup Spotlight I am hoping to dive deeper into this intriguing startup and understand what exactly they are building.

Varda Space Industries (@VardaSpace) is located in El Segundo, in SoCal’s Silicon Beach tech corridor. It is a Series C startup, founded in 2020 by Delian Asparouhov (Founder’s Fund), Trae Stephens (Founders Fund/Anduril), and Will Bruey (former SpaceX avionics engineer). Varda has raised $329M in funding from Caffeinated Capital, Khosla Ventures, and Founders Fund and has a current valuation of $500M.

Subscribe now

Varda is a next-generation cutting-edge Pharma-Space company and it cannot get cooler than this. Varda’s website mentions that they are a microgravity-enabled life sciences company that processes materials in orbit and returns them to Earth. Mind-boggling isn’t it?

I was curious about why we would want to process materials in space when its such an expensive endeavor. The answer lies in the unique benefits of microgravity. The physics of particle movement is dictated by gravity on Earth and it creates the known issues such as sedimentation and this limits the structure of pharmaceutical products that can be made. For example, less dense fluid rises and denser fluid sinks on Earth, but this is different in when there is microgravity. We are longer bound by the physical constraints that gravity imposes, and this gives rise to new pharmaceutical products with unique structures. Scientists have been able to grow more uniform and higher-quality crystals with less defects under the conditions of microgravity.

Varda is actively working on building its capabilities in pharmaceutical production in space and delivering it back to Earth. It’s W-Series is the first-ever commercial satellite and reentry vehicle built specifically for the return of materials from orbit. W series will be launched into orbit by commercial launch vehicles. It’s then a self-sustaining vehicle with a formulation platform and a reentry platform to safely bring the drugs back to Earth.

In May 2026, Varda and United Therapeutics entered a partnership to develop novel formulations for rare pulmonary disease, on Varda’s spacecraft. Despite the exciting science there are definitely challenges ahead for Varda. Commercialization is one of them. Scalability of drugs is a critical factor in determining if a manufacturing process is a viable one and with the exorbitant cost of space flights, and the challenges of manufacturing in low square footage space, there are a million other operational challenges that need to be addressed. But I am optimistic that is endeavor will pave the way for many more. To quote futurist Arthur C. Clarke, "Any sufficiently advanced technology is indistinguishable from magic." and I, for one, can’t wait for the show to begin.

Share

Share Divy's Dispatch

Leave a comment

Hey everyone,

Starting this week, I’ll be sharing a specially curated weekly roundup of the biggest developments across the biotech and pharma startup ecosystem, from who’s hiring and who’s raising funding to key grant deadlines, accelerator applications, product launches, and other opportunities you won’t want to miss.

Keep an eye on your inbox for Startup Intel This Week : your go-to weekly briefing on the news and opportunities!

I’d love to hear your feedback as this series evolves. If there’s something you’d like to see more of (or less of), please let me know. Your suggestions will help make this as valuable and relevant as possible.

Here’s to building, innovating, and staying curious!

Divya

»What’s the Buzz

This week, something has caught everyone’s attention, and had us pause for a second. Stanford University successfully used a generative AI model called Evo 2 to design 16 functional, completely new viruses. These AI-generated viruses are bacteriophages that exclusively target and destroy E. coli bacteria.

» New Deals

• Tarsus Pharmaceuticals to Acquire Alkeus Pharmaceuticals : Alkeus Pharmaceuticals is a private biotech focused on retinal disease. Alkeus's lead program, gildeuretinol (ALK-001), is an investigational once-daily oral therapy in development for Stargardt disease.

• Nuclear medicine leader Curium announced an $8 billion buyout of Lantheus on August 3.

» Get Funded

• FDA Cooperative Agreements (U01): The FDA’s CDER/CBER is funding up to $2.2M per award for projects utilizing digital health technologies (like wearables and sensors) within clinical drug trials.

  • Application Deadline: August 20, 2026

• MassChallenge Breakthrough Bio Accelerator: A prestigious program operated by MassChallenge Boston in partnership with the Israel Innovation Authority, designed to help post-Seed to Series A biotech and MedTech startups navigate international pilot validation.

  • Application Deadline: August 18, 2026

» Trends & Market Analysis

Large pharma companies are racing to acquire precision medicine assets such as radiopharmaceuticals. This was kickstarted by Novartis, which successfully commercialized Lutathera and Pluvicto. Pluvicto is about to hit $5 billion in peak sales.

• Bristol Myers Squibb (BMS) acquiring RayzeBio for $4.1 billion.

• Eli Lilly buying Point Biopharma for $1.4 billion.

• AstraZeneca acquiring Fusion Pharmaceuticals for $2.4 billion.

» Now Hiring

• Strand Therapeutics is looking for a Senior / Principal Associate Scientist, Advanced Cell Engineering (Boston) & a Scientist / Senior Scientist, Process Sciences and Innovation (Boston)

• Peptilogics is looking for a Director, Investor Relations & Communications (Pittsburgh, Pennsylvania, United States, Remote)

• Synolo Therapeutics (launched by RA ventures) is looking for a Senior Scientist/Principal Scientist, Protein Science (Cambridge)

• Output Bio (YC company) is looking for a Member of the Technical Staff, Biological Data (New York HQ, Bay Area)

Anthropic announced its internal preclinical drug discovery program on June 30, 2026, during its "The Briefing: AI for Science" event in San Francisco. During this event, Anthropic’s life sciences head, Eric Kauderer-Abrams mentioned that it will focus its program on developing drugs for neglected diseases which traditional biopharmas have historically written off as non-viable. But why does this category of “neglected diseases” even exist? The economics of pharmaceutical development deincentivizes the investment of capital into any disease that does not present itself as a viable market opportunity. Drug development is known to be a very capital-intensive, high-risk/ high-reward endeavor. Getting a drug to the market is a slow process, which typically takes 12 years to accomplish. The pharmaceutical industry will only take on the risk of developing a drug if there is a potential partnership or a substantial market opportunity. One of the greatest failures of this system is that countries with the greatest disease burden and conditions which primarily affect populations in underdeveloped, poorer regions will not have the benefit of pharma companies developing drugs for diseases which are prevalent in those regions. Anthropic’s focus on neglected diseases may be a real game-changer then. It’s ability to speed up the timeline of these development efforts, might be able to change the unit economics of the whole process of drug development. It will be Anthropic’s moment to showcase how it is capable of rewriting the system.

But realistically speaking, how can a tech company, which is just building its life science capabilities, compete with pharma giants with decades of proprietary data, talent pool and deep domain expertise?

We believe in the power of tight feedback loops, and there’s no substitute for having our own experiences alongside you all in the trenches trying to develop drugs.”

— Eric Kauderer-Abrams, Head of Life Sciences, Anthropic

Anthropic wants to solve the hardest problems themselves, in-house with tight feedback loops, getting deep into the problems, and challenging themselves to develop better solutions. But biology is not software. To translate insilico candidates to the real drugs that get regulatory approvals, there needs to be strong wet-lab capabilities as biology remains an experimental science governed by physical systems, complex cell signaling, and unpredictable living organisms. Anthropic has already begun hiring life science experimental scientists to close the talent gap. But the real question is - will it be enough to get to the finish line?

Sometime over the last few months, we knew it was time to wind down the startup we had worked so hard to build over the last 2 years. The signals were getting clearer. Conversations with prospects revolved around whether the offering was substantial enough to be better than other options they had at that time. There was a lot of resistance to change, intention to stick to familiar products, and tough conversations. All of these eventually led to the phase of fatigue - where you are mentally too spent to take it forward and experiment further. There were really good lessons i learnt in about startups during the course of building which gave me a good understanding of what one should do and not do when building something from scratch. Here’s what I now know :

1. Establish a path to cash flow as soon as you can.

   All businesses have one goal : to have a net positive cash flow to be able to sustain yourself as a business. Conversations with propects and their opinions on how good the product is not useful, unless they put money down on the table. That is the actual act of saying - i see that your product can bring me $$ worth of value and I am ready for it. Unless that happens, any pilot program, compliments, conversations are not very useful. The real litmus test for the product is when a person is willing to part with their money to get hold of your product. That should tell you that you are solving a real problem for them. Instead of checking if a product is solving a “hair on fire problem” by reading reddit pages, talking to people, check if the person who is buying it thinks if its worth it by paying for it. Conversations without testing with payment will just drain your time and energy.

2. Speed is your advantage - hack your way to be unbeatable at speed. Perfection can follow later.

   One of the major advantages you have as a startup is your ability to be agile. Be fast and furious. Make quick decisions and find fast and innovative ways to test your ideas such as building a “hacky” prototype which is will probably take a larger organization longer to get done because of the different levels of decision making that they are entwined in. If you let speed slip away - you will find yourself being outcompeted by every other startup and large organization. Maximise speed for your advantage. But remember - speed doesn’t mean building a perfect feature or a product fast. It means, getting something that could be a diluted version which works just enough to get the basic job done. Keep it high level to get the product out quickly to users’ hands.

3. Prioritize where you want dedicate your time and energy. Delegate if you can.

   Startups can be overwhelming. There is always a lot to do, which will be more than what any small team can handle. At any one time, you might be handling software development, establishing your presence on different social media channels, creating content, talking to users, giving demos, preparing pitch decks etc. The work is not going to go away. The only way to deal with it is to prioritize your workload strategically. A hard pill for perfectionists to swallow but “good enough” should be the goal rather than spending hours on perfecting slide decks or planning for demos. Because if you are like me, you’d probably spend hours preparing for these discussions because you don’t want to seem underprepared. Familiarize yourself with the AI tools available, and find the best way to use them effectively.

4. Set your expectations right. Stay balanced.

   Startups can feel both like a sprint and a marathon. You are racing towards every milestone, but you also have to conserve energy at times to go the distance. It can get overwhelming especially when you have other commitments. Emotions will be intense - both high and low. Getting demo requests and positive feedback from a prospect will feel exhilarating, but as soon as things slow down, and there’s radio silence from prospects when you approach the finish line, it can be absolutely frustrating. Most of us cannot face the outcomes with emotional detachment because of how invested we are in our journey. I found that maintaining emotional balance and not reacting overly to either positive or negative events to be one of the biggest challenges of startup life. If you can focus on doing the work alone with zero expectations on the outcome - that will be most wonderful thing.

   After all the work and the wild rollercoaster of a journey, I’m still so glad I took the leap.I went beyond simply ideating, to establishing a product, gathering users, worked on marketing, creating content and dipping my toes in so many different areas. A startup teaches you more about business than textbooks can teach. The journey may be short-lived but the lessons will last a lifetime. I believe that every failure is a seed for success. The only real rule is to keep going!

   I’ll be sharing more takeaways from this journey in my next few posts.

What is Quantum Computing ?

Quantum computing can be described as an alternatively engineered system for information processing and manipulation which utilises the underlying principles of quantum mechanics such as entanglement and superposition. This is made possible with the smallest information unit, “qubit” which is comparable to “bit” in classical computing. While a bit stores information as either 0 or 1 (binary form), the qubit can be represented by a wave function which is a superposition of 2 different energy states with a combination of the <0> and <1> state. One of the analogies that helped me understand this was a coin flip. When the coin is being flipped, you can imagine it being in neither state but a superposition of both the states with a probability of being in any one state. Entanglement of qubits is basically when 2 qubits’ quantum states are correlated. Entanglement seems to be the way in which qubits “talk” to each other, and they can be achieved by logic gates which makes the qubits intertwine into entanglement and correlate.

What are the Different Modalities of Quantum Computing ?

They are six main quantum technologies that exist today which are fundamentally different in the way they are physically realized.

1. Superconducting Qubits

These are electrical current carrying circuits with a special element called a Josephson junction and are operated at ultra-low temperatures. They can be considered to be a superconducting loop that acts like an “artificial atom” and a qubit on microchips. Companies such as IBM and Google are building their quantum technologies based on this. The advantages of this include its fast gate speeds, 99.5% fidelity and its easy manufacturability given it requires similar microfabrication techniques to ones that already exist. Disadvantages include a short coherence time, and not all qubits are exactly the same with some differences due to manufacturing variability. Moreover, they require energy intensive cooling. The connectivity of qubits is also limited by the grid.The dominant error correction mechanism for superconducting qubits seem to be surface codes which is basically a check on neighboring qubits. That aside, I also read about an interesting new breakthrough by University of Princeton researchers in improving the coherence time of superconducting qubits to 1 millisecond by making changes in the material of construction of these structures (utilizing tantalum (Ta) as a base layer and sapphire as a substrate). This could be an interesting direction of development for this particular modality, given that the paper indicates that this simple material stack can be extended easily from a wafer scale to large scale systems. Some of the questions I had while reading this was, in superconducting qubits, it was mentioned that the losses in the devices are dominated by two-level systems (TLSs). I was not very clear what these Two-Level Systems are and how they contribute to the losses in the superconductors. This is something I have to do some further reading on!

2. Photonic Qubits

Photonic qubits use individual light particles to encode and process information through polarization, path and squeezed state. Squeezed state for photons is something I am still trying to wrap my head around. Some of the significant advantages for photonic qubits are that they can operate at room temperature and do not require extensive cooling to ultra-low temperatures, they do not undergo decoherence in flight, and they are very well suited for high speed communication across optical fiber due to low interaction with environment and high transmission fidelity. Companies like Xanadu and PsiQuantum are focusing on photonic qubits. It’s interesting how PsiQuantum’s approach involves generating large cluster states of entangled photons and measuring them as a gate operation. Their approach is reliant on having very efficient photon sources, and low-loss fiber optics and ultra fast detectors and their direction in improving fault tolerance involves building silicon photonic fabrication to produce chips with thousands of waveguides to create photonic networks but this has not been proven beyond prototypes to extend to large scale fault-tolerant quantum computers. There are also challenges for photonic systems to become a long running cloud-accessible photonic device since there have been issues with stability of source brightness and maintaining the indistinguishability between photons (non-identical photons will create mismatch errors) in long running devices. My opinion is that photonic qubits have a significant advantage due to their ability to not decohere so easily, and they don’t have to be isolated from the environment to maintain their qubit. Therefore, there will be less information loss. But the technology still has a long way to go in terms of being applicable for the FTQC future that we are envisioning with large scale fault tolerant quantum computers.

3. Neutral Atoms

Neutral atom quantum processors use individual atoms which are held in place by tightly focused laser beams, known as optical tweezers. Qubit states are typically encoded into different levels of atomic ground states. By selectively exciting some atoms to Rydberg states (a highly excited state of an atom) , dipole-dipole interactions and Van Der Waals forces of interactions get activated, which can act as gates while preserving coherence. Measurement is done by state-selective fluorescence imaging. Neutral atoms show two-qubit fidelities in the range of 97-98% in experiments with gate speeds of a few microseconds. Although 97%-98% seems large, these errors can compound quickly in a computation.Neutral atoms definitely have the advantages of being very stable, and having longer coherence times in controlled environments and being scalable because large arrays of 2D atoms can be easily generated. Moreover, neutral atoms would have minimal interactions reducing unwanted coupling. The company Atom Computing achieved 1000 atoms loaded into static arrays in 2023 signifying development since this allows for high qubit density without microfabricated circuits. Another company Pasqal achieved 1024 atoms with 0.5 % defect rate arranged in microarrays in April 2026. The more interesting and recent update from this field actually belongs to Atom Computing with their paper titled “ Quantum error correction with the toric code” from June 2026. This paper mentions that they have developed a 1024 atom array which is defect free, using advanced cryogenic shielding, and dual trap lasers. This is exciting for the field of neutral atom quantum computing because this makes a significant progress towards the building large, scalable FTQCs.

Future direction of this field seems to involve designing better connectivity between qubits such as dynamic atom movement although this itself presents other significant challenges such as time for movement. I am curious to find out more about how the movement can be achieved without other challenges such as collisions and energy loss.

4. Silicon Spin Qubits

Silicon spin qubits store information in spin orientation of individual electrons or nuclei in confined in quantum dots or near donor atoms embedded in silicon. The fabrication method is very similar to preexisting methods developed in the semiconductor industry making it significantly easier to scale up. 98% two qubit gate fidelities were achieved in 2019, and 99% was achieved in 2024 with CMOS foundry technology. Previous issues of spin and nuclear noise were improved with isotope purification techniques and improved device engineering. Intel created Tunnels Fall chip with 12 qubits in 2023 with such as extreme ultraviolet lithography (EUV) and advanced materials processing techniques using the 300mm CMOS manufacturing line which will enable it to scale to large scale fabrication easily. Silicon spin qubits’s main advantage lies in its manufacturability, and its small footprint. There seems to be also interest in cointegrating classical and quantum logic on a chip as well.

5. Topological Qubits

Topological qubits are actually quasiparticles that are their own antiparticle. This was one of the most challenging concepts to wrap my head around. These are known as Majorana-based qubits and Microsoft is working on fabricating these qubits.The main advantage of these seem to be a better solution for error correction without requiring multiple physical qubits to form logical qubits for error correction which is a significant challenge that impacts development of quantum systems. These are still in the early stage of development and are not ready yet for large scale applications.

6. Trapped Ion Qubits

Trapped ions qubits are another key modality that exists in quantum computing. These are mainly Barium or Ytterbium ions that are levitated above a chip by electrostatic forces and held in a vacuum. Qubits states would be two sub-level energy levels within the ion.Since they are derived from natural atoms, they have limited variability unlike the “artificial atoms” in superconducting circuits. Gate operations are carried out by laser beams to manipulate their vibrational modes.

Advantages include the highest gate fidelities demonstrated (2 qubit - above 99.9%), high precision gate operations,long coherence and perfectly identical qubits, and all-to-all connectivity. There are also disadvantages with this technology,including slow gate speeds in the order of microseconds,which is several orders of magnitude slower than superconducting solid state qubit platforms. Moreover, trapped ions require a very extensive laser infrastructure with careful design which could be technically demanding to execute. Companies which are leading in this technology are IonQ and Quantinuum. IonQ’s Forte device has 36 qubits, 0.4% 2 qubit gate error rate as reported on their website.They have the highest coherence time and lowest gate error rate of any quantum technology at present. IonQ has also made one the most significant breakthroughs in quantum technology with its recent paper on QCCD devices and walking cat architecture. This is a major milestone in achieving error correction rates that will allow quantum technology to transition from NISQ to FTQCs in the coming years.

The Walking Cat Architecture by IonQ

The paper on the walking cat architecture was complex to understand initially, but as I persisted, I started to find it fascinating as I began to appreciate the elegance of the design.The paper serves as a guide on how trapped ion technology could actually be implemented to become a Fault Tolerant Quantum Computer, which is the goal of all the gate quantum technologies currently in development that I reviewed earlier. The paper also presents an estimate that the Hamiltonian simulation of the Heisenberg model on 100 sites to be executed within a month with 10000 qubits. Just for comparison, I researched on how long a classical computer might take to execute this without massive approximations and I arrived at the conclusion that this might be one of the problems that cannot be solved, and known as computationally intractable on classical computers. From impossible to within a month opens up a world of possibilities in terms of the types of problems that can potentially be solved across many industries and the impact it has on the future got me excited. I recalled an earlier conversation this year with a founder/ researcher who is working on long-range weather predictions. This definitely got me thinking about the different computational bottlenecks this breakthrough could help address.

The premise of the architecture is a QCCD device which allows ions to move in 2D space. The chip is designed with modularity in mind, with specific sites within the chip performing very specific roles. I thought of it like little factories or offices in a bigger city which is the chip itself. The FTQC architecture hierarchy as explained by the paper (Application → Compiler→ Logical Architecture→ Microarchitecture → Device) provides a clear blueprint for how quantum program trails down to actual ion movements in the chip with error correction mechanisms in place. The details of the engineering was mind-blowing to say the least.

5 main areas exist within the QCCD device. Memory blocks where logical qubits are stored and they are cycled in a carousel sort of movement during which the qubits are checked through sensors. A beacon qubit stations itself next to the data qubit and it calls for replacement if a data qubit disappears.

Another modular unit is the cat factory where cat states are produced. These are entangled states which measure qubits. Bell factory acts like a bridge to connect two cat states together so that joint measurement can exist between them even if they are on different parts of the chip and physically separated. Magic factory is where magic states are produced that can be used to implement T logic gates (which seemed to be a more specialised logic gate compared to Clifford logic gates which are the usual CNOT/ Pauli / Hadamard/ S gates). The paper indicates that producing the H states needed for this used to be more challenging using a process called distillation that is more complex and tedious than the methodology proposed in this paper. Apart from these, there are also local and global reservoirs to replace lost qubits and qubit factories to produce more qubits. There also seems to be a need to balance the need of qubits without running out in the global reservoir and supplying too many qubits which will cause the reloading zones to consume too much resources. The paper then provides a logical instruction set for implementation. What struck me as really interesting was the inclusion of a DMX/ DMZ measurement. It seems to be an elegantly designed instruction to optimize speed in situations when just getting the answer is critical but not using the information any further. I read up a little further to understand if this is common practice. What I came across was that this seems to be a standard procedure, but it has been implemented as a formal instruction in this architecture. It got me thinking about what are the scenarios that this might be very useful, but this is something I would have to dig deeper and do further research on.

While reading the paper, I did have several other questions that I was pondering about.

1. There seems to be control within each module, such as replacing used up qubits, ensuring the cat states are produced in factories and quantity required planned ahead before demand. But how does the system prioritize resources for one activity vs the other? Does the decoder control the overall resource allocation? If not, is there a reason to why that is so?

2. How does the width of the memory block impact the type of computation capabilities that are available ? Also, if we were to scale this up, are we in any way limited by the (1) width of the block, (2)movement speed of qubits? Are these like lego bricks that assemble together, and can the qubits move across the lego bricks?

Quantum Algorithms

Next, I reviewed a few quantum algorithms. Quantum algorithms are the procedural instructions for quantum computers, and the key differentiator from classical algorithms is that the unique features of quantum mechanics such as superposition, entanglement and interference can be used to explore different approaches to solving complex problems. The computational part of a quantum algorithm (the gates) should be a sequence of unitary operations.

1. QAOA (Quantum Approximate Optimization Algorithm)

QAOA is a hybrid quantum classical algorithm which is a subset of VQE (variational quantum eigensolvers). QAOA is used for solving problems which have multiple options (combinatorial) and they find approximate solutions for problems. As a hybrid algorithm, it potentially finds better approximations which can be fed into classical computers which gets a better outcome than just using just classical computers. It can be understood deeply by looking closer at a few things. I would define a hamiltonian as energy landscape representation of an optimization problem. Let’s call it Hc. The idea is to convert the optimization problem to a pseudo physics problem where the Hc solution is finding the lowest energy state. The Hc alone would not find the solution here, but you need another operation - which is Hm. The mixer operator informs how you should move through the landscape to get towards the solution. The overall function is then an alternating Hc and Hm function which will be solved to arrive at the solution. The equations for QAOA reflect this approach.

2. QAOA-LR (Quantum Approximate Optimization Algorithm with Linear Ramp)

The QAOA-LR is a slightly modified version of the QAOA discussed above. In this variation, instead of starting the parameters , at random numbers you would be starting at a fixed number and going in a definitive increment called a linear ramp.This will speed up the computation.

3. QFT (Quantum Fourier Transform)

QFT is the quantum version of the classical Fourier Transform. It breaks down a quantum state into its periodic pattern. This is achieved through Hadamard and phase shift gates. QFT seems to be particularly applicable to pattern recognition, data extraction and compression.

4. Quantum Phase Estimation

It helps find the exact ground-state energy levels by estimating the eigenvalues of quantum-mechanical Hamiltonians (estimating the eigenvalues means solving the Schrondinger equation and estimating the allowed energy levels). While VQE & QPE are both algorithms to find the lowest energy state, VQE is a hybrid algorithm using classical computing and quantum but QPE uses only Quantum computing.

5. Grover’s Algorithm

Grover’s algorithm is a quantum computing method for unstructured search that provides a quadratic speedup over classical algorithms. I came across the analogy that Grover’s algorithm is like trying to find an exit of a maze. It amplifies the amplitude of the right solution through interference.

6. QC-AFQMC (Quantum-enhanced auxiliary field quantum Monte Carlo)

This algorithm uses the output from a quantum computer to obtain a trial wave function will be fed into classical counterpart to perform highly precise simulations. Some of the applications of this seem to be in molecular properties and chemical reactions.


(7) AQO (Adiabatic Quantum Optimization)

AQO uses an initial Hamiltonian (energy landscape map) (H0) to slowly evolve into the problem Hamiltonian (Hp) with the final state matching closely with the optimized ground state when the evolution is slow enough. It is currently only applicable to optimization problems.

(8) BF-DCQO (Bias-field digitized counterdiabatic quantum optimization)

This fully quantum algorithm tries to find an optimum solution for a combinatorial optimization problem. Counteradiabatic has an extra driving term which allows the solution to be reached faster than adiabatic, without leaving the right solution path. The algorithm runs in iterations, measuring the qubits after each iteration and checks which direction it’s leaning towards, and nudges the solution in the right direction creating a “Bias”.

Potential Use Cases in the Biotech Domain

Having spent a significant time in the biotech industry across R&D and manufacturing, the potential implications of these quantum computing solutions for the industry are immense and this really excites me. But even before I looked into papers published on the use cases, I challenged myself to first lead the exploration first by looking deeper into areas where I have worked before with this lens and identify a potential use case and evaluate the implications.

This is a potential use case from my previous work that I thought might be interesting for quantum algorithms.

From previous work

(A) Cold chain optimization for drug delivery

The snippet (shown above) from my thesis shows a critical concern for pharmaceuticals. Many drugs need to be maintained within stringent temperature ranges and the challenge is to maintain this cold chain till administration without any potential for spoilage. $35B is wasted annually on this problem and I believe this could be alleviated with logistics optimization, which struck me as something interesting that can be tested with Quantum Computing.

1. Traveling Salesman Problem :

This problem becomes quite challenging to solve by classical methods when a large number of cities are taken.I looked up existing papers where this was tackled and came across Srinivasan et. al’s paper. He outlines a methodology where we could potentially pick either cost / distance/ time to optimize and proceed to solve it with QPE by finding the minimum. In his work, it was proposed to solve this problem by encoding costs between cities as phases.

I initially struggled to understand the approach of how the cost between cities was being converted to phases, and how we are mathematically solving it and arriving at the solution. But pondered and tried to understand the math behind it step by step and finally it made more sense to me than before.

1- Distance/cost/time between cities i and j is converted to Matrix A with Φij values. In his work, he takes cost as the parameter to optimize.

2- Convert the Matrix A to Matrix B by encoding the cost as phases. [ B = e^i(Φij) ]

3- Obtain the Unitary Matrix

4- Apply the phase estimation algorithm

5- Apply the quantum search algorithm on the cost database created to find the optimal route.

Based on my initial inspection of the work done in quantum for biotechnology, I decided to split my exploration into two segments. Tier 1 - there are some interesting use cases, with small proof-of-concepts and subsequent papers published. Tier Two- Areas I think might benefit from QC, but do not have any substantial proof-of-concepts.

Tier 1

(B) Protein folding simulation

I found a paper by IONQ “Protein folding with an all-to-all trapped-ion quantum computer” where BF-DCQO algorithm was used to determine protein folding with 12 amino acids by searching for the lowest-energy conformation on a two- or three-dimensional lattice.

Questions : How are we determining the exact mathematical way to “MAP” a problem to Quantum systems? Won’t there be situations where we completely get the encoding wrong - wrong way to solve the problem? How do we prevent this situation?

(C) Calculating ground energy of molecules :

I found this paper : Use VQE to calculate the ground energy of hydrogen molecules on IBM Quantum
Question : How do we pick an Ansatz ?

Tier 2

Here are other use cases that I think might be interesting but potentially need to be broken into smaller problems we can work on with quantum computing first.

1. Bioreactor optimization for product yield - multifactor combinatorial problem

Bioreactor optimization is a multidimensional, non-linear, highly dynamic system which is challenging to optimize and computationally heavy. In an attempt to use quantum computing for this use case with continuous decision variables, and discrete variables, Hernandez-Romero et al (15) provide a possible approach to solve these problems in their review paper and mention that Hybrid- Quantum-Classical algorithms (QAOA) will be useful in this case. After reading the paper, I did have the question of how we can utilize empirical data in this HQC.I was also wondering why reaction kinetics was chosen to be solved by classical solvers - since i also believed that kinetics, is a typical energy minimization problem that falls under the problems quantum solves well.

Mapping chemical reaction mechanisms - Energy minimization problem

2. Drug-Target interactions / drug docking energetics - Energy minimization problem

Quantum computing may help in predicting binding affinities and optimize drug candidates since these are fundamentally based on energy minimization and on quantum interactions at the molecular scale.

3. Optimizing LNP (lipid nanoparticles) formulation - Multifactor combinatorial problem

The current methodology for optimizing LNP formulation is through Design of Experiments, which is carrying out a limited number of experiments with varying parameters, such as ratio of lipid particles, temperature and other few other variables. These experiments are mainly used to understand the interactions of the different variables. This approach might not lead to the optimum solution, and there is a risk of obtaining a “local optima”. Given that quantum computing is useful in optimization of combinatorial problems, this will be an interesting problem to tackle.

Hands on Experimenting

Next I wanted to get hands on executing a code on one of the quantum computing platforms. I used anaconda and accessed IBM’s qiskit by creating an account. Since I was new to this, I used Claude as my coding partner to write the code, but encountered several errors along the way. With a little bit of research, and debugging, I managed to get the code running.

I also found that I could solve a sudoku 2x2 puzzle with Grover’s algorithm.

Final thoughts

This weekend, I spent the time exploring quantum computing deeply. I think of this exploration like a jigsaw puzzle where I was fitting the pieces by reading one piece by one piece, trying to understand what the underlying concept was. I started off reading about technologies, and then real use cases. At some point when I was trying to fit the real use case with the actual logic of encoding computation with quantum mechanics, I kept getting stumped by one missing piece. How does the implementation math work? How do we know how to work through a problem we want to solve with quantum. I focused alot on the gates, the mathematical operations. But although I understood it at a high level, I was still questioning : Why are we doing these complex mathematical operations, how are we arriving at the answers? It hit me then, that for a certain class of problems, because nature naturally tends to seek an optimum energy minimized state, all we are doing here is latching on to that universal truth of physics to help us solve problems and everything became clearer. The whole jigsaw came together and it was simply elegant and fascinating. My understanding of quantum computing evolved as well in the process. Before I read up on quantum computing, I thought it would be an advantage to use quantum computing for every computationally heavy problem. But after this exercise, my understanding has matured to discern problems which are more suited for quantum computing such as optimization problems, energy minimization, simulation of molecules and interactions. With that realization, I must confess that I had a really satisfying “workout”.

Other questions for which I need a little bit more clarity:

• How do ions explore many different possible solutions?

• How is the right option optimized through wave interference?

References

1. https://academic.oup.com/nsr/article/12/8/nwaf246/8165689

2. https://arxiv.org/pdf/2503.14798

3. https://ieeexplore.ieee.org/abstract/document/11421923

4. https://www.youtube.com/@Lukas-Lab/featured

5. Chapter 2 - Quantum computing: vision and challenges : https://www.sciencedirect.com/science/chapter/edited-volume/pii/B9780443290961000088?fr=RR-2&ref=pdf_download&rr=a156ba99c9a94307

6. Review of Distributed Quantum Computing: From single QPU to High Performance Quantum Computing :

https://arxiv.org/abs/2404.01265

7. A Fault-tolerant, Ion-trap-based Architecture for the Quantum Simulation Algorithm : https://people.eecs.berkeley.edu/~kubitron/courses/cs252-S10/projects/reports/project3_report.pdf

8. https://arxiv.org/pdf/2604.07205

9. https://atom-computing.com/wp-content/uploads/2026/06/AtomComputing_QuantumErrorCorrectionwithToricCode.pdf

10. https://www.ionq.com/quantum-systems/forte-enterprise?utm_term=ionq%20forte%20enterprise&utm_campaign=GR+-+Brand+-+Global&utm_source=adwords&utm_medium=ppc&hsa_acc=6464340917&hsa_cam=18671018300&hsa_grp=149338283886&hsa_ad=663875618608&hsa_src=g&hsa_tgt=kwd-2451006822147&hsa_kw=ionq%20forte%20enterprise&hsa_mt=e&hsa_net=adwords&hsa_ver=3&gad_source=1&gad_campaignid=18671018300&gbraid=0AAAAAoctzZszhR4MxeAiMjvc06f14xUSg&gclid=Cj0KCQjw3qLSBhDaARIsAFTiVh6H3k68nI-kCdoefNKYwyJ0BC5PJN6ofjxq3rP3dinGRSQrkQmHOdAaAgVSEALw_wcB

11. https://arxiv.org/pdf/2402.17520

12. https://arxiv.org/pdf/1805.10928

13. https://arxiv.org/pdf/2506.07866

14. https://arxiv.org/pdf/2305.06538

15. https://pubs.acs.org/doi/pdf/10.1021/acs.iecr.4c04624?ref=article_openPDF

The pre-PMF phase is really not for the faint-hearted. One image that comes to my mind when i think about it is : imagine yourself walking through a dense jungle at night with a flash light. You can only see what’s immediately in front of you. Based on that limited knowledge, you take one step forward. You look out for signals. In the jungle, you are probably focusing on the sounds, making sure it’s safe to proceed. You evaluate if it is the right step or you should change your course. This is exactly what you do while you build your startup. Except now, its not jungle sounds that you are focusing on, but rather the market signals. You evaluate the outcomes of every step and decide if you are headed in the right direction.

I am trying to think through this whole idea of figuring out the strategies to achieve PMF as i type this. I am no startup guru. I am just thinking out loud. When you start building, you probably wont have much of an idea on what is right and what is the smart direction to take to tackle the pain point you aim to solve.Many startup advisors recommend doing customer discovery calls. Customer discovery calls are severely limited in offering any substantial value to the founding team in my opinion. We all have gripes with different tools. “ I wish this tool had this feature, I wish it did this better etc”. But how many of us would actually pay for the tool even if those issues were fixed? Let’s go back to first principles. If you are selling a product directly to a consumer, let’s think about his monthly budget first. He/She probably has rent to pay, grocery to buy and other substantial monthly expenses. What would make him allocate a significant portion of his monthly budget to your product? What would the person value so much more than the dollar amount he pays month over month for your product? Is it extra time( in the case of productivity apps)? Better health ( in the case of fitness monitoring)? Is it peace of mind ( in the case of security or insurance products)? Or is it simply status and identity (think luxury goods, or even that premium gym membership)?

The point is, the value exchange has to crystal clear. It has to be so obvious to the customer that parting with their money feels less like a painful transaction and more like a relief. Not “this is nice to have.” Not “I’d probably use this.” But “I genuinely cannot imagine going back to the way things were before this existed.”

This goes back to my initial gripe with customer discovery calls. They have a hidden trap.Most of these calls hover around the topic of all the nice things that you can build. As founders, we value every call, and we of course are so eager to please, and listen to everything the customer says. But, if you can intently steer the conversation away from NICE to MUCH NEEDED, I WILL PAY FOR IT THIS SECOND mode of questioning, it’s going to be the guiding light towards the PMF treasure in the jungle. If it ends up being the former, you should still take a neutral approach to the discussion, because it might fool you into building a bunch of useless features that hardly moves the needle.

China’s New Focus on AI+ Manufacturing

China has been quietly revamping its manufacturing industry to future proof it in the AI era. China’s Ministry of Industry and Information Technology (MIIT) and eight other agencies are working towards a collaborative effort known as “AI+Manufacturing”, which will be rolled out through 2027. This ambitious program aims to revolutionize China’s manufacturing infrastructure with a focus on making them smarter, efficient, and autonomous. Given that China is the world’s largest manufacturing power, with nearly 30% of the world’s manufacturing value, this massive push for AI integration could be the spearhead, that leads the changes in the new industrial revolution and setting the benchmark for how industries worldwide adopt intelligent technologies.

Ambitious Goals for AI Integration in Manufacturing

The transformation that China is envisioning is not a simplistic one but a fully integrated, end-end infrastructure change. The government is focusing on every step of the manufacturing operations, including research and development, production, logistics, supply chain, predictive maintenance, intelligent scheduling, digital twins and machine led quality inspections. Moreover, there is a push to release industrial AI models which can be sector specific and highly specialized to optimize and enhance manufacturing operations. Chinese manufacturers will soon be looking into incorporating agentic AI in order to analyze manufacturing data and act on behalf of humans and make decisions about manufacturing processes. With the increase in real-time data, and emergence of intelligent manufacturing operations with agentic capabilities, there are new avenues to solve age-old problems with China’s manufacturing sector : environmental impact. These systems can track energy consumption and emission data in real-time during production processes and create opportunities for optimization of processes.

A clear example of this comes from a case study of Xiaomi, a global leader in electronics. Xiaomi is making strides in sustainable manufacturing through its smart factory in Yizhuang, Beijing, which produces millions of smartphones annually. Xiaomi has utilized AI effectively to minimize electronic waste and use raw materials more effectively with minimal wastage. Overall, these efforts have enabled the company to improve efficiency and cut material waste by up to 30%.

Driving Forces That Are Fueling the Change

Major tech companies in China are racing to become the key orchestrator for AI agents. Developments by these major players are fueling the agentic revolution in China’s manufacturing industry.

• Alibaba’s Gwen 3.5 & Wukong Platform: Alibaba’s Gwen 3.5 a new AI model (released Feb 2026) designed not just to generate text, but to perform complex tasks independently

• Baidu released an entire suite of OpenClaw-based agents for desktop software, cloud services, mobile tools, and smart home devices.

• Tencent integrated OpenClaw into WeChat via ClawBot, a feature that appears as a simple contact in the chat window. Users send it tasks, it executes them.

Moreover, China is making immense progress in the development of humanoid robots. The startup Agirobot aims to create machines that can collaborate efficiently with humans, further accelerating automation and efficiency in China’s industrial ecosystem.

Clearly, China’s initiatives for AI in manufacturing is well-positioned to serve as a blueprint for the next industrial revolution, shaping how manufacturing industries worldwide adapt to technological innovation in the coming years. However, it will be interesting to observe what risks, safety measures, governance frameworks, and ethical safeguards are implemented to mitigate the uncertainties and unknown risks inherent in the rapid integration of cutting-edge AI technologies.

Subscribe now

250 biotech and pharmaceutical executives from nine countries - Canada, US, France, Germany, Italy, Spain, UK, Brazil, and Switzerland, took part in a study conducted by MIT Technology Review Insights in October 2025.

The results of the study indicated that 75% of respondents are actively planning, piloting or deploying agentic AI initiatives. Some of the use cases that most executives are immediately focused on are regulatory compliance (51%), data standardization (49%), patient support (46%), and market intelligence (46%). The signals are clear. Agentic AI has immense potential to transform workflows in pharmaceutical organizations.

Pharmaceutical organizations present some of the clearest use cases and fastest paths to ROI for agentic AI.

Having worked across both small biotech firms and large pharmaceutical companies, I’ve seen firsthand where friction exists. It’s not a lack of talent or data. If anything, it’s the opposite.

Why Pharma Companies Have Great Use Cases for Agentic AI

Pharma companies operate with large, highly specialized teams generating immense volumes of data at every stage of the R&D lifecycle. But that scale comes at a cost. Teams work in silos and knowledge doesn’t flow as easily as it should. Coordination across functions discovery, translational research, clinical, regulatory is often slow and fragmented.

Workflows remain deeply manual. Critical processes like literature review, data harmonization, protocol design, and regulatory documentation are still stitched together across disconnected tools and systems. And while these systems were built for a different era, they’ve become deeply embedded making them difficult to replace, even when they no longer serve their purpose effectively.

I experienced it first hand when I worked on a research problem once within a large pharma organization. I tried to find relevant previous studies done within the organization to take the idea forward. It was a such a challenge to gather insights from all the studies and side projects that various people had worked on related to the topic I was working on, and gathering and building on top of the insights generated by these studies. The inefficiency of the system appalled me. This was probably months of work - but here it was- being unusable for any further application, with messy data that someone else has to figure out from scratch.

McKinsey’s white paper on Agentic AI for Pharma indicates that 75 to 85 percent of workflows can be enhanced or automated by AI agents across functions such as biostatistics and data management, medical writing and medical affairs, safety and pharma vigilance , analytical chemistry, manufacturing and quality, clinical science, regulatory affairs and compliance.

How Agentic AI Drives Real-Time Decisions and Execution

One of the key advantages of Agentic AI is the ability to process complex unstructured data in real-time and not only provide insights but take decisions and execute on follow up actions. This ability allows for faster decisions, utilize predictive analytics to foresee trends and bottlenecks, optimize for resources, scale production, manage demand fluctuations, reducing downtime and improving reliability.

The High Cost of Innovation: Why Drug Discovery Is Ripe for Agentic AI

Drug development is a costly endeavor even with today’s technological advancements. Studies have estimated that the R&D cost for a new drug ranges from $314 million to $4.46 billion, depending on the therapeutic area, data, and modeling assumptions.Success rates, remain critically low, with approximately 90% of drug candidates failing during clinical development phases. At the earliest stage, computational studies have become fundamental to developing better targets for various diseases. Deep learning models have become more valuable in making accurate predictions of molecular structures and properties, allowing for more accurate assessment of absorption, distribution, metabolism, excretion, and toxicity (ADMET). Agentic AI has the ability to autonomously carry out insilico experiments and ADMET predictions, without downtime and autonomously allowing for running multiple experiments in parallel. This, along with wet-lab data processed with agents, can speed up the entire workflow at least 30%.

The convergence of high costs, complex workflows, fragmented processes, massive data volumes, makes pharmaceutical organizations an ideal proving ground for Agentic AI. To fully realize its potential, Agentic AI must operate reliably, with transparent traceability and well-defined guardrails, ensuring both effectiveness and accountability.

Subscribe now

Last week in my weekly musings, I was going through the series of steps of how I started execution, when my idea for a business was still a tiny thought in my head. How much we have come from a thought to a whole new tangible product! The journey has been nothing short of amazing. One of the first resources i relied on when i knew that I had to take my idea further was YC’s startup school.YC runs one of the world’s most well-recognised, and sought after accelerator program, with each cohort getting 20,000-40,000 applicants. They are really the go-to people for everything startup, having had multiple unicorns as their alumni from these cohorts such as Drop Box, Door Dash, AirBnB etc. The library of resources for those figuring out entrepreneurship is completely insane.

A huge shoutout to the YC team for putting together such valuable content. YC’s startup school has a short video on everything you need to know when you start your company, such as how to come up with a good idea, how to develop a prototype/MVP, how to talk to customers and lots more. What I loved about each video was how succinct, powerful and interesting it was to listen to each of them. I did not listen to them like I listened to a lecture. Entrepreneurship was captivating. I listened to them when I had a long drive, or when I was in the gym, and whenever I had a chance to squeeze in a couple of minutes. The more I listened, the more I was drawn into the world of entrepreneurship. And the more I was drawn in, the more I was focused on execution of my little idea. I realised that ideas were dime a dozen. Execution was everything. I knew I had to have a team to execute an idea well. One of the best tools YC offers is also its Cofounder Meeting Platform where enthusiastic founders can meet others, find those who have complementary skills and work together as cofounders towards a mission. This is the first big step in a Startup Life. Getting a team together. If you already know your team mates from before, this process is easier. You have a history, you know each other’s strengths and weaknesses. But if there is no one from your network who is as passionate about startups as you are, you need to find new avenues such as Cofounder Matching Platforms to help you find the right person.

Looking back now, it’s fascinating how fragile ideas are in the beginning. They exist only as fleeting thoughts. Easy to dismiss, easy to postpone. What resources like YC’s Startup School did for me was simple but powerful: it turned the intimidating concept of “building a startup” into a sequence of small, executable steps.

Suddenly, the path from idea to product didn’t feel abstract anymore.

Subscribe now

A Century of Science is Now at the Verge of Changing

Scientific progress and innovation have been at the core of human civilization since humans first began documenting their discoveries.What started with speculations on mythical creatures and superpowers evolved into scientific laws and frameworks that guide every discovery and invention. The underlying curiosity of humans to understand their environment, use that knowledge to make progress and build a better future, has been the primary driver of science since the start of time.In our relentless pursuit of a higher state of existence, we have created new drugs for diseases previously thought incurable, created technology that makes knowledge accessible instantly and connected people across oceans. Most of the scientific progress in the last 100 years had been about gathering knowledge from existing work, identifying gaps and working on the foundation of some key seminal papers and answering questions beyond the discoveries of those papers. Prior to the LLM-era, most of the knowledge scouting work was done by researchers by evaluating every source and fitting every source’s relevance to the research question. But with the advent of AI, the process of scientific discovery may undergo its most fundamental transformation in over a century.

Autonomous Agents: Ushering in a New Era of Science

In August 2024, Sakana AI announced the introduction of their AI Scientist. A fully automated scientific discovery pipeline including the entire research lifecycle, from generating novel research ideas, writing any necessary code, and executing experiments, to summarizing experimental results, visualizing them, and presenting its findings in a full scientific manuscript. This was the most audacious plan for scientific discovery built for the era of artificial intelligence. While this created a huge buzz in the academic and AI community, initial excitement soon turned into mixed feelings. The paper, “Evaluating Sakana’s AI Scientist for Autonomous Research: Wishful Thinking or an Emerging Reality Towards ‘Artificial Research Intelligence’ (ARI)?” by Beel et.al indicates several key shortcomings, including the lack of robustness in the process and simplistic methods of identifying novelty.

In May 2025, Futurehouse, a San Francisco-based nonprofit research lab primarily funded by former Google CEO Eric Schmidt, announced the launch of its platform- “Superintelligent AI Agents for Scientific Discovery”. It features four specialized agents. Crow for general queries, Falcon for deep reviews, Owl for prior work checks, and Phoenix for chemistry planning. Each of them have been designed for high precision and transparent reasoning. With access to open-access papers, specialized databases, and scalable API/web interfaces, Futurehouse’s platform aims to automate workflows, analyze literature, identify gaps, evaluate methods, and accelerate discovery, including chemistry-specific tasks like hit identification and reaction prediction. Futurehouse then unified the agents into a system known as “Robin”. Robin was utilised to identify ripasudil, a Rho-kinase (ROCK) inhibitor clinically used to treat glaucoma, as a novel therapeutic candidate for dry age-related macular degeneration (dAMD), a leading cause of irreversible blindness. The process of determination of the therapeutic drug was systematic and co-ordinated with a series of workflows that was a hybrid loop between physical experimentation and computational analysis.

• Robin used its agent Crow to do a comprehensive literature review to come up with a hypothesis. The hypothesis was : Enhancing retinal pigment epithelium (RPE) phagocytosis could provide therapeutic benefit for dAMD.

• Falcon was utilized to identify a set of molecules that could be good potential candidates.

• Laboratory experiments were carried out.

• Finch performed data analysis on experimental data obtained in the lab.

• Robin then proposed an RNA-sequencing experiment to investigate how Y-27632 increased phagocytosis in retinal pigment epithelium (RPE) cells.

• After the experiment was conducted, the AI agent Finch analyzed the data and found that Y-27632 upregulated ABCA1, a key lipid efflux pump in RPE cells, helping explain the observed effect.

• Using results from the first round of drug testing, Robin proposed a new set of candidate molecules.

• Subsequent experiments identified ripasudil, an existing ophthalmic drug, as the most promising candidate for treating dry age-related macular degeneration (dAMD).

Researchers from other organizations have started to test with FutureHouse agents with some agreeing that it is better than general agents for identifying and synthesizing literature specific to a research area.

However, there is still a lot of gap that needs to be filled in order to truly be the AI scientist that can co-invent with humans.

Challenges In Building an Autonomous Scientist

One of the key challenges of an autonomous scientist is the biases that might exist in the AI model. Although humans do have their own confirmation bias- the process of rigorous review and criticism from peers can identify those biases and reduce the risk of personal biases affecting scientific decision making. However, in the case of an autonomous system acting on its own, there needs to be a strong strategy for implementing and monitoring those bias as researchers may have minimal involvement in ensuring those biases do not affect the scientific decision making. This can lead to an echo-chamber effect with AI systems reinforcing each other’s biases rather than identifying them as mentioned by Resnik et.al in the article, “Autonomous artificial intelligence, scientific research, and human values”. As such the process needs to be clear and transparent with clear explanations of every step. But with a black box AI system, reasoning behind every minor decision point might not clear leading to its own set of challenges.

Another challenge is whether LLMs can engage in deception and malicious activity. Anthropic’s new AI model Claude 4 Opus, which can work autonomously on tasks for extended periods of time, demonstrated concerning behaviors such as attempting to preserve its existence and even blackmailing an engineer in a simulated scenario to avoid being replaced, during testing. Concerning behaviors like these can pose significant security risks if left unchecked. Autonomous AI scientists will need robust safety guardrails to ensure that their actions remain aligned with human intent and do not compromise research integrity or safety.

A Balanced Approach to Scientific Agents

Artificial intelligence will undoubtedly become an integral part of our lives. Organizations must adopt a balanced approach that maximizes the possibilities of AI while ensuring that scientific rigor, safety, and ethical standards are not compromised.

What’s writing without a personal connection? Nothing! That’s why I decided to dedicate a small corner of this space on the internet (My Substack) to also share about my entrepreneurial journey, and my life around it. On most days I write about the latest AI innovation. Other days, I hope to give you a window seat to what my life looks like. I’ve lived in 3 different countries, lived and worked across many cities in North America. Science and innovation has always been my passion, and hence I dedicated myself to biomedical engineering and drug development. Together with that, I developed a deep, undying love for entrepreneurship. I am one of those people who comes up with new business ideas and shares it with friends at dinner parties. I think of how every opportunity can be turned into something bigger and better. I enjoy the process of creating value from nothing, especially building products that push the boundaries of what is possible. But my life experiences have also moulded my personality to be better suited for the game. I have moved to a lot of new places, started from scratch multiple times, and established myself in each place. These experiences teach resilience, the ability to accept and embrace change and uncertainty - which are very important skills as one learns to navigate through the roller coaster of startup life. But despite all the passion for entrepreneurship, life kept me busy with corporate life and I did not have such a grand idea that moved me enough for me to take serious action.

I had a spark of wisdom one day, just like a lightning bolt and I realised it was THE IDEA that I had to work on. Interestingly, I got to work the day this thought came to my head. What followed was an interesting run of events… (TO BE CONTINUED)

A significant deal happened in the tech industry recently and this might be the moment a major AI tool for research took a serious step for positioning itself as a trustworthy research infrastructure for enterprises.

On January 29, Perplexity signed a $750 million deal with Microsoft, giving it access to frontier AI models from OpenAI, Anthropic, and other leading developers.

For researchers, this is a game-changer. Perplexity now runs on dedicated, enterprise-grade compute in Azure, which directly improves speed, reliability, and depth of source verification. More compute means it can pull from more papers at once, cross-check claims across multiple sources, and return answers with consistent citations, especially critical when dealing with long, technical documents like journal articles, patents, and regulatory filings.

Microsoft’s enterprise security and compliance also mean Perplexity can now be safely adopted in corporate R&D, pharma, and consulting environments, where AI tools were previously blocked due to privacy concerns.

And it’s not just Microsoft. In May 2025, Wiley, one of the world’s largest and most trusted academic publishers partnered with Perplexity to bring publisher-verified, peer-reviewed scholarly content directly into the platform. Wiley is now Perplexity’s first education partner, giving students, educators, and researchers at subscribing institutions access to licensed collections in fields like nursing, business, and engineering all inside Perplexity, complete with proper attribution and citations.

This integration addresses one of the biggest challenges in academic AI: misinformation and unverifiable answers. Instead of relying solely on web scraping or AI-generated content, users can now combine authoritative Wiley resources with live web information, with every claim properly sourced. Pilot programs are already running at Texas A&M and Texas State, with UK institutions expected to join soon.

Why this is signficant is mainly because of the higher degree of institutional trust that Perplexity is bringing to its products. Perplexity is positioning itself as an AI tool that universities and administrators can openly adopt, rather than restrict, which is a shift from AI tools operating outside academic guardrails toward systems built directly into trusted scholarly workflows. For researchers, that means AI can finally accelerate discovery without compromising rigor.

With that context in mind, it’s a good time to step back and look at where Perplexity came from in the first place.

Founded in 2022 by former members of OpenAI, Meta, Quora, Bing, and Databricks, Perplexity wasn’t built to replace Google or casual chatbots. Its founders were motivated to solve the limitations of traditional search engines, which he believed were designed for the web of the early 2000s rather than the AI era. His goal was to move beyond a “list of blue links” and create an “answer engine” that provides direct, accurate, and synthesized answers to user queries, backed by citations.

Instead of generating text first and attaching sources later, Perplexity retrieves and evaluates information first, then generates an answer grounded in verified citations. This design choice immediately made it appealing to a specific audience: researchers, analysts, graduate students, and knowledge workers who are those people who need traceable, defendable information and not just fast responses.

In research, the main challenge has not been finding information. The more critical challenge that was unaddressed was deciding what to trust, what to prioritize, and how to synthesize it into valuable insight. Perplexity sits at that junction: not replacing reading or thinking, but reducing the groundwork of collecting that information and presenting it a quickly analysable format.

Let’s look at who actually uses Perplexity today, and how it fits into real research workflows because that’s the market Perplexity aims to capture. Perplexity has 22 to 30 million active users, with 780 million queries per month of 2026. It’s looks huge, but compare it to chatgpt whose numbers are more insane. ChatGPT has 800 million weekly active users currently and Over 2 billion daily queries.

Perplexity’s impact in the research world is growing and in the right direction. For now, however, users have quickly pointed out that several limitations still remain.

• Such as oversimplifying methods, statistical nuances, or experimental caveats or providing surface level answers for highly niche technical papers.

• Perplexity struggles with highly multi-step reasoning, and

• Hallucinations and inaccurate citations occurs now and then.

With the Microsoft deal and the Wiley partnership, Perplexity has officially made the move from experimental AI to serious research infrastructure. It’s getting geared towards getting direct, sourced, actionable answers, and synthesizing information in a way that saves researchers hours of groundwork.

For PhD students, grad students, and industry analysts, that means you can quickly map a field, compare studies, and gather trustworthy references, all in one place. It’s designed for the new AI era of research, not the old web-of-links model.

But let’s be realistic. Perplexity isn’t perfect. It can oversimplify complex methods, miss statistical nuances, struggle with multi-step reasoning, and occasionally hallucinate or miscite. For highly technical or niche topics, the original papers are still irreplaceable.

The smart way to use it? Treat Perplexity as your intelligent research assistant. Perfect for literature triage, initial analysis, and synthesis but never a substitute for critical thinking or deep expertise.

The main requirements for an autonomous system is an end-end agentic system that will utilise telco network models and AI agents that can not only talk but also use network simulation tools to validate actions. NVIDIA has seeded the AI transformation in telecom by inaugurating a Nemotron-based large telco model (LTM) along with a guide for building reasoning agents for network operations, and NVIDIA’s Blueprints for the optimized energy savings and network configuration to create a multi-agent system for an autonomous, self-managing telecommunications operations.

The Nemotron LTM is an open model. One of the key benefits includes the clear visibility that is given into its training process and underlying datasets, allowing them to deploy it securely within their own infrastructure. Running the model within their own infrastructure means operators aren’t dependent on external cloud environments to power their AI systems. They can design, deploy, and manage agents directly inside the network core. At the same time, the model can be continuously adapted using proprietary operational data, allowing telecom teams to strengthen domain-specific reasoning while keeping full ownership of their data, compliance posture, and security controls.

Not only that, NVIDIA’s NeMo skills pipeline can enable operators to fine tune their reasoning model to develop highly specialized agents. These agents can operate as network engineers. Each incident category, resolution procedures can also be documented to translate into a a reasoning framework that to make these agents more effective at making decisions. The result is a network that doesn’t just respond but it evaluates, simulates, adapts, and improves.

And this time, the ambition isn’t efficiency alone. It’s self-managing, intent-driven infrastructure.

Agentic systems are being hailed as the future of AI but behind the buzzwords lies a surprisingly structured choreography of planning, reasoning, and action. An agent system is simply an autonomously acting workflow that can understand requirements, and accomplish tasks without needing any special API access because of its ability to work like a human who is using a computer but does not get tired until the task is done.

Agentic systems are usually made of a number of agents working in unison. Each agent will have a very specialized task that it is fine-tuned for. The starting point for an agentic execution is initiated by the user, who provides a context and the requirements. The user’s task will be to explain the task and the goals clearly such that the agentic system has the necessary information to create an optimal output that meets all the requirements.

Workflow of Agentic Systems

The workflow is fundamentally similar to what happens in a real world scenario in an office. The lead agent analyzes the given task and breaks it down into smaller simpler steps. Then, the lead agent will assign each of the individual agent who is specialized in a performing a certain type of task to carry out each substep. Each specialized agent would be having the background necessary to execute the task based on previous experiences (memory) and the expertise gained in that specific area of focus. The agents have access to internet data which is available to them, as well as organizational internal data which they have been provided access to. Then, there is collaboration between the agents to systematically execute the task at hand. There will also be iterations based on feedback of the user, and questions from the agent and the responses provided by the user.

An example of an agentic system could be one where there is :

• Planner agent which plans the method of execution

• Research agent which gathers the data necessary

• Analysis agent which analyzes the gathered data.

A New Workplace

As more organizations focus on amplifying their AI capabilities, agentic AI has taken the center stage as of today. Companies are looking for the right execution layer instead of just incorporating a knowledge layer. Agentic AI is a very viable and cost-effective solution for many repetitive workflows which are time-consuming, and resource-heavy. As organizations evolve with AI, so will the role of employees. Instead being an executor, the employee has to be the perfect orchestrator who is able to give the right context, and optimize the output of these agentic systems.

While we extensively used AI over the last few years to summarize dense text, explain concepts without clicking a dozen links, search for “new” ideas, write emails which were politically correct yet sharp, paraphrase passages that needed to sound more polished to go beyond our own capabilities of writing, AI has been quietly evolving to not just be our thought partner, but to become something more of an Assistant whom we can entrust tasks to.

Massive improvements in short and long term memory structures have enhanced these AI systems to understand context much better, across sessions, and adapt to the users short term and long term goals, and act in alignment with accomplishing those goals, while consistently delivering better results over time.

Just a few days ago, Perplexity launched “ Perplexity Computer”. Perplexity Computer already sounds intriguing. It’s a name that feels unexpected, almost out of nowhere, with a subtle sense of nostalgia. Perplexity Computer is a digital worker, that can not only answer questions but can do tasks. Perplexity Computer is a system that creates and executes entire workflows over even long periods of time. What’s even more mind boggling is that each computer has a number of “sub-agents” which will perform individual tasks such as research, document generation, and coding etc. These agents are automatically delegated, and coordinate the assigned work among each other. You can also run many Perplexity Computers in parallel. Imagine having an army of workers to do all your tasks and you as the leader orchestrating your army. This is precisely where AI is today. Autonomous agents doing tasks without human interference.

Just a 3 weeks ago, OpenClaw created a ripple in the tech industry with the launch of its autonomous agents. It caused a sudden surge in demand and unexpected shortage in Apple’s hardware suppy chang as AI developers snapped up the 2024 M4 Mac Mini. OpenClaw is a personal AI assistant that runs locally on your mahine and connects to messaging apps you already use such as iMessage, Slack, Signal or Discord and 20 others. Created by Austrian developer Peter Steinberger, its an open source AI tool that allows an agent to interact with websites, and other apps like a human would. Scrolling, clicking , typing and reading a screen, without needing an API or special integrations, and performs the task that you assign it. It could be something like booking a flight for you, or negotiating a car deal or doing some specific research and generating a document. It does this by using a vision model to interpret whats on your screen and then follows by deciding what action to take using a large language model. It then takes the action, observes the response, and this continues until the desired goal is reached. Since it doesn’t have predefined workflows, it can adapt to UI changes, and handle ambiguity very well. This tied with persistent memory makes it a very compelling tool for handling tasks autonomously.

So what does the future hold with these digital workers performing tasks while we, humans sleep? While they don’t have a circadian rhythm, nor a need to take coffee breaks, the real challenge and execution bottleneck now is : how good is the context provided for the agents to deliver a solid output and no longer the effort factor. The quality of the output now depends on the quality of the intent, constraints, memory, and direction we provide. Moreover, with everyone deploying agents now, the real question is not whether should we use agents. It has become - which agent is actually better in accomplishing my tasks like an owner and not an intern.