AI/ML Platform Engineer in Boston. I build production AI systems, streaming data platforms, and the MLOps glue between them.
A control plane for the LLM lifecycle — evalstack gates changes in CI, tracelens observes them in production, and routerai governs cost by routing each request to the cheapest model that clears a quality bar — alongside RAG-grounded NL→SQL and a self-hostable data platform for feature engineering. All MIT-licensed, all demo-able locally.
Open-source LLM evaluation framework — Python SDK + CI plugin, LLM-as-judge rubrics, regression detection, side-by-side run + event diff. Gates prompt/model changes in CI before they ship. Braintrust alternative.
Alpha · Demo liveNatural-language analytics for founders + PMs — schema-aware, RAG-grounded, sqlglot-safety-gated SQL generation against DuckDB. EXPLAIN-based cost preview, single-page chat UI, key-free demo mode.
Production LLM observability — one decorator to instrument, then a failure-path breakdown (rate_limit / timeout / connection_error) plus latency + cost per model. FastAPI collector, pluggable storage (SQLite or ClickHouse), p50/p95/p99 dashboard. Auto-detects OpenAI + Anthropic response shapes.
Alpha · Demo liveCost-aware LLM routing gateway — routes each request to the cheapest model that clears a quality bar, with shadow-eval and savings metrics. The "govern" control plane that completes the trilogy: evalstack gates, tracelens observes, routerai governs.
Real-time feature store you can deploy in a day — declarative YAML feature defs drive a recompute-on-write pipeline into DuckDB (offline, Parquet-exportable) and a pluggable online store (in-memory or Redis). Feast/Tecton alternative for teams without weeks to spend on deployment.
MVP · Working end-to-endEnd-to-end open-source data platform template — the feature-engineering backbone under ML/AI systems. Postgres → Debezium / snapshot CDC → Iceberg-style Parquet → dbt (bronze/silver/gold) → DuckDB queries. Single make demo runs the full path.
