ClaimsBench: Evaluating AI Agents Across a Complete Insurance Claim
How ClaimsBench evaluates AI agents across evidence gathering, adjudication, revision, and closure in complete insurance claims.
Projects
Research projects, experiments, and papers on reinforcement learning, agents, and evaluation.
Building a Legal AI Application With SimpleMem, AutoResearch, and Legal Skills
A local Legal AI application that combines document upload, SimpleMem conversation memory, legal workflow skills, and retrieval optimization for contract and legal question answering.
Recursive Failure Archaeology: RLM for Agent Failure Diagnosis
A recursive investigation pipeline that diagnoses agent failures by segmenting long trajectories, generating hypotheses, pruning evidence, verifying controls, and comparing against full-context GPT-5.5.
Trace2Evolve: A Karpathy-Style AutoResearch Harness for Customer-Support Agents
A research-style AutoResearch harness for improving tool-using customer-support agents using benchmark traces, heldout tau2 evaluations, synthetic pressure tests, and reliability-gated promotion.
Teaching an LLM to Explore: Reinforcement Learning for Document Navigation