At Percepta we build decision making systems for the hardest operational problems at a business. Running them at scale gets expensive fast, so @shohamshitrit, @AlbertYZuo and team built infra to make private evals for our customers and then hillclimb the pareto frontier of cost vs accuracy.
In our latest example, we built a system that reads the full clinical timelines of 100k patients every day and picks the next best actions for each one.
The task is prohibitively expensive even with Gemini Flash models, so we trained a distilled 0.8B open model that beat every frontier model we tested at >20x lower cost.
Michael (@michaelvartikar) on our team recently built Kaizen - a tool that helps us build private evals for our customers and then hillclimb agents against those evals 🧵
1/ We built a compiler that turns arbitrary programs into transformer weights. You give it C code, it produces a standard transformer that executes that program step by step.
If I had a nickel for every time Payton Pritchard checked in for a half court buzzer beater in the NBA Finals…I’d have…two nickels, which isn’t a lot, but it’s weird that it’s happened twice