On ClauseIQ, I built a contract-review agent backed by SQLAlchemy and PostgreSQL, with hybrid retrieval and deterministic risk checks. Its frozen benchmark reached 95% Recall@5 on an 80-query holdout, and evidence-grounded evaluations check outputs before they’re trusted.
I also built ProofPatch, a program-repair system that accepts fixes only when its verification gates pass; its reproduced benchmark achieved 100% certified precision with 0% false certification. Across my projects, I focus on reproducible evaluation, auditable evidence, and systems that abstain when the results aren’t reliable.
