The lead is current production client work, a working content automation solution, and applied AI case studies. Public repos make the hands-on work inspectable.
Owned the enterprise generative-AI portfolio for a global Medical Affairs organization: a production RAG platform serving ~2,000 users across 50+ markets, an ~$8M annual GenAI and digital portfolio, governed under GxP and 21 CFR Part 11.
Built and governed retrieval-grounded generation for regulated Medical Affairs workflows: hybrid retrieval, ranking strategies, evaluation rubrics, grounding and drift review, human-in-the-loop gates, LLM-as-a-judge, and audit traces. FastAPI / retrieval / AWS. Confirmed hands-on by a technical peer.
Led Celgene Medical Affairs IT workstreams across 20+ platforms, 1,500+ users, and 50+ markets in 18 months. Delivered $7M in operating efficiency and $4.3M in annual cost avoidance.
Live AI agent for an anonymous law practice with multi-matter shared-folder ingestion into Postgres, resumable scans, deduplication, bounded OCR, passage-level lexical retrieval, conversation and matter memory, retrieval evaluation, recovery gates, and repository-side backup operations.
Working independent stack covering signal intake, duplicate checks, brand routing, structured authoring with human review, deterministic rendering, and channel operations with readback. An anonymous client architecture remains in the specification phase.
Precision-first pipeline with URL normalization and deduplication, ambiguity abstention, deterministic rules, structured model judgments, evidence grounding, audit trails, and a separate evaluation set reporting 0.900 precision and 0.866 F0.5.
Clean-room synthetic prototype that validates exterior-maintenance photo packets, runs an observation-only provider path, fails closed on weak evidence, writes immutable run artifacts, and keeps the human decision separate.
Semantic memory MCP server (spec 2025-11-25) with hybrid FTS5 keyword plus vector recall, running locally with no per-call token spend.
Live product that turns discovery-call transcripts into client-ready proposals, SOWs, estimates, and HIPAA BAAs, with deterministic WBS estimation, PHI gates, and audit logs.
Independent case work. The system abstains when product identity cannot be proven, and every accepted result stays grounded in the supplied evidence.
Independent synthetic prototype only. No customer data, no payment or work-order mutation, and no claim of production adoption.
Checkpointed agent runs with eval gates, and a typed MCP server scaffold with Zod-validated tool inputs.
Reproducible ML-evaluation research on how metric choices and weak bootstrap assumptions reverse model rankings, and a recompute of a labor-market capability framework against the 2026 frontier.
Explore the complete portfolio on the main site.