A small optimization demo, followed by examples from machine learning, AI for chemistry,
agent systems, process development, and electrochemistry.
↳Try it yourself
Optimize a process you can't see.
A small two-parameter optimization demo. Click points on the surface, or let the optimizer
suggest the next experiment.
0experiments
·best so far
Click the surface to run an experiment.
The optimizer uses a Gaussian process surrogate with Expected Improvement, similar to
obsidian,
and runs in the browser.
Selected examples
01AI & Modeling · Biologics
A neural-ODE architecture for cell culture
GARNET began as a way to model complex small-molecule reaction mechanisms and now models cell culture and fermentation. I wrote it end to end in PyTorch and torchdiffeq.
Challenge
Fit the coupled behavior of roughly 15 metabolite and in-process measurements against roughly 20 process parameters, from very few experiments.
Approach
Dropout neural networks integrated through an ODE solver, with custom regularization, constraints, and numerics for stiff systems. Handles bolus glucose and media additions without stopping and restarting integration, which is what makes feed optimization possible.
Results
Usable models from as few as 30 experiments, which changed how teams design bioreactor campaigns and cut process-characterization workload by up to half. Feed-strategy optimization improved titer 15% and harvest-day viability 20% on a pipeline program.
Deployment
Extended to dynamic pH control, cutting bioreactor control-strategy experiments from close to 200 to as few as 50. Also deployed as a hybrid-model digital shadow supporting deviation analysis and real-time decisions at a manufacturing site.
02AI & Agents
Agent systems scientists actually trust
LLM agents that do real work for scientists, and the evaluation that keeps them honest.
Challenge
Getting agents past demo quality is an accuracy problem, not a plumbing problem. Scientists stop using a tool the first time it is confidently wrong.
Approach
Tool-use harnesses over our own models, so agents can train models, run simulations, and generate experiment designs. For scientific summarization I built an adversarial loop that grades every draft against ten criteria and forces the generator to revise until it clears the bar.
Evaluation
Curated test suites of increasing complexity that score whether generated instrument commands actually execute, used to show how much system-prompt structure drives reliability. Retrieval is tested against fixed prompt sets built around cases where near-identical names change the correct answer.
Results
In production for hundreds of scientists, alongside deterministic connectors that make lab notebook and LIMS data reliably readable by models and applications.
03AI for Chemistry
Teaching optimization to read chemistry
Bayesian optimization is good at numbers and bad at molecules. This work closes that gap.
Challenge
Process optimization has to choose a catalyst, a solvent, and a base alongside the temperature and the loading. The standard answer is one-hot encoding, which means the optimizer learns nothing from the fact that two ligands are chemically similar, and campaigns get very expensive very quickly.
Approach
End-to-end deep learning from featurization through acquisition. SMILES transformer and graph embeddings compressed by deep-kernel learning into a small set of latent dimensions the Gaussian process can actually use, so chemical identity and numeric process parameters are searched together.
Results
Twice the rate of selecting the best ligand compared to one-hot encoding, across 32 diverse ligands on a real 300-point experimental dataset. Manuscript in preparation.
04AI & Modeling
Bayesian optimization for process experiments
Practical tools and training for Bayesian optimization in process-development experiments, and the vendor partnership that carried it further than my own code could.
Challenge
Adapt Bayesian optimization for day-to-day process-development use.
Approach
Founded Algorithmic Process Optimization, built and taught the tooling, and published obsidian, the department's first open-source codebase.
Strategy
I set the approach my department still follows: keep the advanced algorithms in-house where the science is, and pair them with a vendor product for scale. Merck is a pharma company, not a software company, and I would rather my team be expert in the algorithms than maintaining an interface. It has since been applied to hybrid modeling, lab automation control, and high-throughput chemistry plate design.
Result
160 active users across small-molecule and biologics programs, running roughly 50 optimizations a year; individual campaigns met their targets using 25 to 75% fewer wet-lab resources. The vendor grew from two people to more than ten over the same period.
A COVID-19 antiviral needed a process redesign on a compressed timeline.
Challenge
Develop a robust API-forming reaction and control strategy in four weeks.
Approach
Ran hundreds of high-throughput reactions, built kinetic models, and used simulation analysis to support a complete process redesign.
Result
More than a 10x wider stability window, an order-of-magnitude lower impurity burden, higher yield, and improved robustness for the manufacturing route.
Most synthetic electrochemistry is reported at small scale. This work carried it through to kilogram scale.
Challenge
Take a clean, selective electrochemical oxidation from grams to kilograms without losing control of it.
Approach
Built a workflow for scaling electrosynthesis, ran a kilo-scale thioether-to-sulfone oxidation, and engineered a spinning three-dimensional anode for a multigram Hofmann rearrangement.
Result
Among the first kilogram-scale synthetic electrochemistry in the industry. OPRD Outstanding Publication of the Year, 2022.