At Feast, I implemented two protobuf-based feature-server endpoints, wrote an end-to-end lifecycle test, and blocked UDF unpickling in infrastructure payloads. At Docling, I fixed an ODT hyperlink-extraction regression across three parsing paths and added edge-case tests.
I built ARGUS, an 11-agent financial due-diligence system using Google ADK. It routes narrow questions to targeted agents, reducing model calls from 15–20 to 2–3 compared with the full workflow, and includes claim verification, guardrails, and a human-approval gate.
I also built Arbiter, a cost-sensitive fraud decision engine evaluated on IEEE-CIS transactions. A calibrated XGBoost–LightGBM ensemble achieved 0.5597 PR-AUC; I used temporal validation, added fail-closed serving and an audit log, and tested the system with pytest.
