Engineering Intelligence With Purpose

Thomas To

Founding Fullstack AI Engineer

Canventa Life Sciences

Engineering Intelligence With Purpose

Thomas To

Founding Fullstack AI Engineer

Canventa Life Sciences

Thomas To-40 Under 40 USA 2026

AI becomes far more interesting when it is built for problems where reliability cannot be a nice-to-have in discovering, developing, and manufacturing medicines. Thomas To has spent his career working in precisely that space, where biochemical engineering, scientific research and software intersect, first exploring computational protein modelling at UC Davis and later moving deeper into software and AI at Genentech and Canventa Life Sciences. His work has since expanded from scientific computation into production systems, combining machine learning, RAG agents, data infrastructure and open-source contributions to solve problems that once depended heavily on manual processes. Along the way, his projects have generated $63.2 million in modelled manufacturing savings, digitised more than 5,000 laboratory records with 95%+ accuracy and reduced decision cycles from more than three hours to under 10 minutes.

The thinking behind those systems is equally important, shaped by a belief that AI in scientific environments must earn trust through validation, reproducibility and measurable outcomes. TradeFlock Magazine speaks with Thomas about the experiences, principles and engineering choices shaping his approach to AI, software and life sciences.

Which turning points moved you from biochemical research toward software and AI?

The journey did not begin with a decision to work in AI. It emerged through problems that kept pointing me toward computation. Four years of protein engineering research across UC Davis labs with Nandi, McDonald, Wan, and Siegel showed me the ceiling of manual wet-lab iteration. I built models in pyRosetta and PyMOL, validated them against wet-lab results, projected $63.2 million in manufacturing savings and produced three published protein variants. The deeper lesson was that computation could multiply laboratory work rather than replace it.

Genentech made that lesson practical. At the Vacaville site, I watched scientists still track production manually. Automating those pipelines cut processing time from weeks to minutes and pulled me toward software as the leverage point. Moving into software engineering at Genentech South San Francisco made the shift permanent, especially after building tools for five-plus scientific teams.

What is the life-sciences AI industry still getting wrong about scientific data?

The industry has good insight into the empirical realities of scientific data. However, with the commercialisation of language model providers, agentic AI must also support established and novel processes, in addition to cross-functional domain subject matter expertise. While the general public and other non-technical stakeholders are applying conversational models to scientific and regulated data, much more engineering goes beyond a chat interface. I made that case publicly in International Business Times, arguing that AI should support scientific decision-making rather than replace empirical biology.

Language models are not limited to other AI modalities and methods applied to the life-sciences AI industry. Protein engineering taught the same lesson earlier, when computational designs still required wet lab validation before they could be trusted.

Scientific AI needs domain-specific models, rigorous evaluation and empirical validation. Treating AI output as fact rather than a hypothesis is where these systems fail in practice.

What principle became non-negotiable in the way you build technology?

Validate against held-out ground truth before anyone trusts a system in production.

The principle formed at the UC Davis lab bench. Computational designs built with pyRosetta and PyMOL predicted protein function well, yet some still failed wet lab testing before three variants ultimately held up and were published. Computation without empirical validation can produce confident, wrong answers.

I carry that standard into every AI system I build. The Arctic-TILT pipeline shipped only after held-out evaluation confirmed 95%+ accuracy on laboratory records feeding compliance and finance processes. The RAG guardrails built for Canventa’s revenue model follow the same logic because outputs reaching non-technical stakeholders need validation inside the pipeline. Every project since that lab bench has been built to catch the gap before a stakeholder does.

"Computation without empirical validation can produce confident, wrong answers."

POWER PROFILE CARD

Age: 28

Country: USA

Industry/Sector: Biotechnology

Years in Leadership: 6

Leadership Philosophy: Leading by example to support the next generation.

Motivation: Building technology that solves meaningful problems and leaves people, systems and the field better than he found them.

When reliability matters more than speed, how do you make the trade-off?

Accuracy and reliability take priority when working with AI systems that require compliance or support scientific decision-making. Speed and cost come after you meet the reliability bar.

Canventa gave me a practical test. More than 5,000 handwritten laboratory records needed digitising, but an off-the-shelf model couldn’t meet the fidelity requirements of regulated data. I fine-tuned Snowflake Arctic-TILT against systematically held-out validation data rather than shipping a faster pipeline. The slower approach reached 95%+ accuracy, saved 1,000+ hours of manual transcription and preserved the audit trail compliance teams depended on.

A demand forecasting model reinforced this by preventing a $2 million inventory stockout and $200,000 in material waste. An accurate system that runs slowly can be more valuable than a fast system nobody trusts.

What has open-source work taught you about designing production AI systems?

Open-source work taught me to design for hardware and infrastructure I do not control. Contributions across OpenXLA, JAX, Kubernetes, and OpenFold3 exposed me to environments where production systems cannot depend on a single preferred stack.

While working through Dr Noah Flynn’s Build AI Drug Discovery Pipelines book, and trying to design and open-source my own zero-cost end-to-end architecture for AI Drug Discovery, I learned that production AI systems require teams of software engineers across data, backend, machine learning, frontend, and deployment.

Understanding where and how data flows through AI systems will guide whether to use open-source solutions from third-party organisations or individuals as internal or external dependencies to meet the business’s functional and nonfunctional needs.