LexLoop
A contextual bandit learns which of three LLM agents solves a legal case best, rewarded via an LLM-as-a-judge rubric.
Problem
Three LLM agents can work on the same legal case, but not equally well. The project lets a learner decide which agent to use for which case.
Approach and architecture
A contextual bandit picks one of three LLM agents for each case. Its reward comes from an LLM-as-a-judge rubric. The environment is built with Gymnasium (RL environments), model calls go through LiteLLM, and tests run with pytest.
My role
I built the environment, the run logging and the Apertus integration, and I am responsible for documentation and QA.
Status and results
In progress. The repository is private.