Independent AI research lab
Deepform Labs builds instruments for understanding intelligent systems: what they learn, how they behave, and where their apparent capabilities break down.
Modern AI systems can look stable from a distance and behave unpredictably up close. A single number hides the contours that matter.
We develop methods that expose those contours, measuring systems across tasks, conditions, and time instead of treating intelligence as a fixed point.
Our work starts where standard evaluation stops: beneath the aggregate score, inside the trajectory, and across the changing conditions of a deployed system.
Measurement for AI systems.
Instruments that tell you what a model actually does, not what a leaderboard says it does. Current work examines measurement error, agent trajectories, serving conditions, and fine-tuning regressions.
View program ↗Separate meaningful model gains from noise, variance, and evaluation artifacts.
Study agent trajectories and intermediate decisions, not only final task completion.
Measure how quantization, serving load, and system context reshape behavior.
Map capability tradeoffs and hidden regressions introduced during fine-tuning.
Build tests that capture behavior at the level where it changes.
Vary prompts, models, environments, and system conditions deliberately.
Follow the path between input and outcome, not just the endpoints.
Turn the observed structure into tools people can use and verify.
Measure the system in front of us, not the story told about it.
Preserve complexity where simplifying it would hide the answer.
Make claims proportional to what the experiment can actually support.
Collaborations · Research · Early access
mohith@deepformlabs.com↗