Bryan Falkowski
Papers, projects and notes on evaluating AI and building with coding agents. Most of it has code you can run.
Checking whether an eval run worked before trusting its score.
Measuring and correcting LLM judge scores, with a customer service politeness example.
Whether LLM test agents refuse when the target element is missing or ambiguous.
How consistently LLM judges reproduce scores across repeated runs.
In progress
Decides what evidence a pull request needs, checks it, and learns its rules from bugs that got through. Built by a Building 58 room.
CodeChecks whether an eval run is valid before its score is trusted.
CodeA deployable system for building, testing and improving AI agents. Room B uses it as a new user would.
CodeCompares a purpose-built decision model with a Claude agent on the same browser test.
CodeA containerized service for running LLM evaluation jobs, with a console and Kubernetes deployment.
CodeShort story and short film
A short story inspired by Player Piano, then a 3D kitchen world and a first-draft film made in one day with AI tools.
Read the storyFilms
The people edition and the robot edition, each in 3D and realistic versions.
Watch