Inspiration
I work in a lab, and a lot of people I know work in labs too. One of the biggest bottlenecks we run into is learning to use specialized equipment. You can understand the science behind an experiment and still need someone to show you which settings to use, how to interpret a reading, and when you can trust the number on the screen. Onboarding often means finding the one person who knows that instrument and catching them when they have time.
That experience inspired EVOLVE, our Scientific Software Factory. We wanted a researcher to be able to describe what they’re trying to measure and get an app that makes the instrument’s limits clear as they work.
What it does
EVOLVE turns a measurement request into a tested analysis app for a specific lab instrument. In our demo, a researcher asks, “What’s the concentration of this mystery vitamin B12 sample?” The first reading on a simulated spectrophotometer falls outside its calibration range, so the app explains why it can’t give a reliable estimate.
Once the lab lead approves diluting the sample, EVOLVE builds and tests a new version of the app while the researcher prepares the vial. The new version accounts for the dilution. Anyone reviewing the estimate can trace it back to the readings, dilution volumes, calibration, and code that produced it.
How we built it
We started with a registry that describes what each instrument can do. Given a measurement request, Codex generates the Python analysis, the steps the instrument should follow, and the screen where the researcher inspects the result. The factory checks those steps against the instrument’s capabilities and tests the generated Python offline against independent reference cases before releasing it.
We first used the factory to build two apps for electron paramagnetic resonance spectroscopy. One measures the width of a spectral line. The other analyzes how signal intensity changes with temperature. The B12 experiment let us test whether the same factory could handle a different instrument and revise an app midway through a run.
Challenges we ran into
One of the hardest parts was deciding what should count as a successful test. For the undiluted B12 sample, the correct behavior was to withhold the concentration estimate. When we added dilution support, we had to verify two things at once. The revised app had to recover the original concentration at two different dilutions, and it still had to recognize the original reading as out of range.
Accomplishments that we're proud of
Our favorite moment in the demo is when the first reading comes back out of range. It’s exactly the kind of moment when a new lab member might stop and look for help. EVOLVE explains the limit, leaves the decision to dilute with the lab lead, and builds the software needed to continue the experiment. A reviewer can inspect both versions and see exactly what changed.
What we learned
Working on EVOLVE made me realize how much of onboarding is about learning judgment. An experienced lab member knows when to trust a reading and when to prepare the sample again. To make that knowledge useful in software, we had to turn “that reading looks too high” into a specific calibration limit, a test case, and an explanation a new researcher could act on.
What's next for Scientific Software Factory
We want to take EVOLVE from simulated instruments to a real lab bench. I’d start with one instrument in my lab and ask an experienced user and a new lab member to run the same workflow. Where does the new user still need help? Which checks does the experienced user make almost without thinking? Their answers would guide what we build next.
Built With
- 8090
- codex