Building Transfyr

The business model of science is broken

Science is the most transformative engine for human progress. It yields the medicines, materials, and infrastructure that make modern life possible. Yet, the process of turning a discovery into an impactful and scalable product relies on tools that have not kept pace. Knowledge stays locked in local labs, communicated effectively through 1:1 apprenticeship, while incomplete protocols and lab notebooks thwart reproducibility. What we choose to share publicly is even more limited: papers that show only the final success, while completely losing the record of failed attempts, expert instinct, and subtle variation that made it possible.

We are living through a period of extraordinary acceleration in individual technologies, yet our methods for communicating scientific execution have remained remarkably stagnant. This growing gap cripples momentum: experiments fail to reproduce, scale-up falters, money runs out, teams disband, and the next generation finds it harder to get the resources they need to even get started, because it’s hard to invest in a leaky bucket. Fundamentally, what we see is a crisis in the distribution of science. The product of science incorporates tacit knowledge, and knowledge is hard and expensive to transfer. Business 101: when the product of an enterprise can’t be distributed, the enterprise fails.

We founded Transfyr to tackle this challenge head-on. We’ve spent our careers investing in and building organizations at the cutting edge of innovation, taking on moonshot challenges and getting frustrated at the waste and loss we saw when great innovations failed to achieve liftoff. The challenge that demands our attention now isn’t any specific disease or crisis, but what is missing from all of science: the infrastructure layer that allows scientists to understand and transfer the knowledge that keeps the engine running.

As we may see

The vision of what scientists should be able to capture in the lab and draw insight from has existed since the time when computers weighed 30 tons and photographs were still manually developed in chemical baths.

In 1945, Vannevar Bush, the architect of American science policy that shaped the 20th century, imagined a visionary future where technology would store, index, and surface the sum of human knowledge to extend our intellectual reach for the advancement of science. He lamented that “Professionally our methods of transmitting and reviewing the results of research are generations old and by now are totally inadequate for their purpose.” We can only handle, capture, transmit, and draw insight from a fraction of what is actually happening, and our abilities to learn and make connections are limited by the tools of data capture, storage, and retrieval, locked into the structures of how computer architectures were built, not the ways we think and learn. Unfortunately, several generations later, we still live with similar challenges and the limitations of lab notebooks and PDFs of journal articles.

Bush predicted digital archives at everyone’s fingertips, speech to text, thinking machines, and—most vividly—a headmounted “cyclops camera” that could capture much more than what a scientist thought to write down in his notebook: “As the scientist of the future moves about the laboratory or the field, every time he looks at something worthy of the record, he trips the shutter and in it goes, without even an audible click. Is this all fantastic?”

80 years later, every prediction Bush made about the tools that would help us see, capture, and learn from the world has come true, in nearly every valuable domain of business and technology, except for the protagonist of Bush’s vision: the scientist at the bench.

The missing layer

Observability is an essential part of high stakes business. Factories are filled with cameras that monitor for quality control at every step. Competitive sports referees use instant replay to confirm critical decisions and top athletes go back to the film room with their coaches to review and improve their performance. Software companies deploy continuous monitoring to confirm production systems are healthy and quickly trace any failure back to its source.

Scientists don’t have any of that. Even for experiments that can cost millions of dollars, use priceless, irreplaceable, samples, or hold the key to new cures, scientists don’t have a reliable way to look back and audit the record. Their best bet is to attempt to decipher the results of their experiments. When those results are unexpected or inconclusive, when a collaborator can’t replicate them, or when a manufacturing run fails, the scientist inevitably ends up in endless meetings rehashing what could have happened from people’s memory and incomplete records. A high stakes decision on what to do next in the absence of trustworthy data leaves the scientist with a sinking feeling, needing to take a leap of faith on a guess of what went wrong.

At best, when science works, we’re left with a lossy, biased record of what happened. The methods section of a journal article is a snapshot, an outline of what worked at a given moment in time, making for a fragile breakthrough. Our methods to communicate and transmit research learnings remain “totally inadequate for their purpose,” leaving irreproducible results and stalled innovation.

AI has entered the picture, promising to massively accelerate science and support the most consequential decisions in drug discovery. But AI can’t learn from what was never written down. Without observability into the real-world work happening in a lab, the frontier of what AI can do remains jagged where it matters most.

Likewise, lab automation promises to relieve scientists of manual work and accelerate the enterprise of discovery and development, but translating a new protocol to automated systems is slow, painstaking work. All the unwritten details of what scientists actually do with their hands need to be translated into executable code. The work of science, it turns out, is hard to automate when there is no data to learn from.

Without observability, science is less usable, reproducible, scalable, and translatable. Without observability, AI and automation can’t make a difference. That’s why we’re building Transfyr.

Unapologetically ambitious and pragmatic: a North star for the journey ahead

The narrative around AI for science has fallen victim to polarization. AI will either eradicate all disease or unleash the next pandemic. Robots will either do all the jobs or are incapable of doing nuanced scientific work. Our experience has allowed us to embrace the unstable equilibrium.

Renee is a moonshot hunter who has led some of the most ambitious scientific programs in the country, but has seen how mundane failures of human coordination stymie progress. Anna Marie was trained as a spreadsheet junkie, but knows what’s possible when brilliant scientists meet impossible problems. We’re both unapologetically ambitious - we’re excited about a world where science moves at the speed of software. We’re also unapologetically pragmatic - we’re just as happy to step into the muck of boring, annoying problems if it means enabling a scientist to do their best work and be “top of license” (a term Renee taught Anna Marie, meaning spending time doing the things one is uniquely qualified to do).

We believe robotics will generalize, commoditize, and transform enormous swathes of manual work, and we’re collaborating with multiple robotics teams to enable robotics to be useful to science. However, we’ve also seen the widespread failure of the promise of automated cloud labs, where big vision was whittled down to a few reliable operations, or worse, where the equipment collects dust, turned “on” only when an investor or customer is walking through the lab.

We believe that we can fundamentally change the way we implement work across the entire enterprise of science AND make a meaningful difference in the everyday life of a scientist at the bench.

We believe there is enormous potential for good in the acceleration and widespread distribution of scientific knowledge but are also thoughtful about building in safety at the outset given the inherently dual-use potential of what we’re building.

These shouldn’t be either-or decisions. To achieve the vision of AI transforming science, we must do both.

Transfyr is working to build a lossless system of record for scientific execution, so the next scientist, model, or robot can learn faster. Our focus this past year has been on capturing scientific metadata. This is easier said than done, because science is full of nuanced, almost imperceptibly tiny details that really matter, and many actions can only be judged in the context of prior or subsequent actions. Improving our tools to passively capture and draw insights from the scientific metadata will be a continuous journey, but it’s required to enable reproducibility, tech transfer, and ultimately the acceleration of frontier science across contexts.

While we aim for a less monstrous term than “Cyclops Camera,” we’re bringing Vannever Bush’s vision for a frictionless, lossless record of scientific observation to life. We deploy tailored sensor stacks integrated with models trained on scientific execution into labs across frontier research, clinical diagnostics, and beyond, creating a high-fidelity record that enables root cause analysis, supercharges training and workforce development, and guides the development of automation and AI.

We’ve built our portfolio of customers and projects across four strategic domains that matter for the future of science:

1. Frontier AI and Robotics: We didn’t get autonomous vehicles by uploading maps to AI models. A map is an extremely lossy representation of the road; Waymo needed to install sensors on cars and learn from data from real-world driving conditions to train their autonomous fleet. In the same way, a Nature paper is an extremely lossy representation of science. Transfyr is creating the real-world data needed to build autonomous scientists. We’ve partnered with the top frontier AI labs to build the data and, importantly, the biosecurity toolkit and real-world evals needed to set the standard for building a highly capable, and safe, frontier AI stack. We’re also excited to be partnering with the DAMP Lab at Boston University on their $20M NSF / Genesis Mission program aimed to create a network of autonomous labs to support the development of a strong US bioengineering infrastructure.
2. Workforce Development: Teaching and training often gets relegated to “soft skills,” but when a pharma CEO is negotiating with state governors about expanding their manufacturing footprint to their state, the first question they’re asking is whether they can get the skilled operators they need. We’re proud to be supporting the incredible team at BioBuilder on a nearly $1M program funded by the Massachusetts Life Sciences Center to advance AI-assisted workforce development.
3. Scientific Operations: Across both frontier academic research and routinized scientific operations, observability is crucial for understanding and optimizing processes. We are collaborating with researchers in some of the most innovative and prolific academic centers in the world to uplift their deeply skilled scientific operators and accelerate their scientific processes. We also collaborate with scaling diagnostics and operations facilities to help teams understand inter-operator variability, finicky protocol steps, and opportunities for optimization.
4. Tech Transfer: Moving a breakthrough from the bench to a contract manufacturing site or clinical partner is where millions of dollars and months of patent exclusivity disappear. By replacing static SOPs and lossy handoffs with a continuous, high-fidelity system to capture scientific execution and understand what matters, we seek to ensure that science transfers seamlessly, scaling innovations out of the lab and into the real world without losing months to troubleshooting.

We’re excited to kill the endless meetings, the expensive troubleshooting, and the blind decision making that today are an inevitable consequence of the lack of observability. Even our earliest prototype systems were able to surface critical insights. At scale, observability offers a foundation of process understanding that plugs the leaks that prevent science from being reproducible in new contexts. When we can routinely distribute science across the globe, we crush the bottleneck between science and impact.

We’re lucky to work with the best of the best

We are tackling some of the biggest challenges in physical AI, applied to the world’s most impactful domain. We have assembled an unbelievably talented and mission-driven team that is taking on these ambitious challenges with pragmatic focus. Our team has built labs on almost every continent, deployed some of the largest scaled automation systems in science, authored the leading DNA foundation models, and aced our (nearly) impossible technical challenges (most people who finish it get 10-15%, if you can get >90% in our holdout set, you’ll go straight to a final round interview!). They’re also weirdly good cooks.

We’re also lucky to have a brilliant group of advisors from across the many domains that we touch. Their support helps to guide the complex union of frontier science with frontier AI on our platform:

• Chris Ré - Stanford professor, and one of the leading researchers in efficient model architectures (FlashAttention, state space models, Evo 2)
• David Baker - Nobel-winning professor at the cutting edge of computational protein design, whose lab has also been pioneering observability solutions for decades
• Kevin Weil - the storied product leader who was most recently the Chief Product Officer and Head of Science at OpenAI
• Steve Quake - the leading biophysicist and former Head of Science at CZI
• Ken Frazier - the former CEO of Merck
• Jakob Uszkoreit - the CEO of Inceptive, which builds AI generated RNA therapeutics, and author of the foundational AI paper on transformer architecture: “Attention is All You Need”

We’re also joined by investors who share our vision for science that can move at the speed of software. We’re grateful to Hemant Taneja at General Catalyst for leading our $25M seed round and to the incredible investors at Breakout, Factory, Lux, SV Angel, Neo, Underscore, MVP Ventures, and Lyda Hill for their support and partnership.

We’ve built our team with four key values in mind:

• Team, not turf: We’re working on a uniquely hard, cross-disciplinary challenge. While AI will enable some one-person companies, this is not one of them. Anna Marie still orders the office snacks from Costco and Renee’s not afraid to wield a pipette. There’s no room for politics or turf wars here - there’s more than enough work to go around!
• Scientific revolutions, not science projects: while there’s a lot of frontier research we’ll tackle at Transfyr, it only matters if it reaches the real world. Our team is motivated to see our innovations make a difference to customers.
• Bias toward action: We give a shit. This is important work and we operate with urgency.
• Top of license: We are building infrastructure to allow scientists to spend more time at top of license and we want to enable our team to do this as well - we seek out ways to allow people to spend as much of their time as possible in the places where they thrive.

We’re hiring across our technical teams at all levels. Apply on our web site or shoot us an email directly (recruiting@transfyr.ai). Please note that we get far more applications than we can review with care and so if Transfyr is the place you want to be, the best way to stand out is to get a warm introduction from someone we trust. We also recently launched a research fellowship (more details at fellowship.transfyr.ai) for exceptional researchers who overlap with the many areas where Transfyr is working at the frontier.

The business model of science is broken, but it doesn’t have to be. We shouldn’t be satisfied with excuses (“science is just hard”) or empty promises. The future depends on what we can see. Once we started paying attention, we noticed all the little things that were avoidable and unacceptable. Now we can’t unsee it, and we hope you can’t either. Welcome to Transfyr.