Data and AI engineer. I build systems where correctness is the feature — production data pipelines, applied AI tools, and the seams between them.
About me
An offline AI classroom for low-resource environments. Runs on a single laptop — no internet, no telemetry — built around Gemma 4 and Ollama, structured as a curriculum, not a chat box.
An AI-native two-stage decision tool for residential entitlement work. Site screening plus a structured pre-acquisition memo, with Claude API output guards against hallucinated specificity.
A reference implementation of an idempotent enterprise integration pipeline. Five-branch deduplication, circuit-breaker DLQ routing, ambiguous-outcome resolution, replay with full audit trail.
By day, I build and maintain production data integration pipelines at INTO University Partnerships — Salesforce and SQL Server flows, reconciliation, exception handling, audit trails — on the integration team supporting 19 U.S. university partners.
Outside of that, I’m building Beacon for the Kaggle Gemma 4 hackathon, iterating on LandClear with development partners, and maintaining the open-source Integration Pipeline Console alongside a three-part series in preparation on enterprise integration reliability.
An idempotent write path for multi-system integration: business-level keys, a lightweight dedup store, and a three-layer pipeline that makes side-effect correctness queryable and auditable.
How pipelines survive partial failure and uncertain delivery state without amplifying errors — retry classification, circuit breaking, dead-letter routing, and the design of ambiguous-outcome queues.
Raising throughput from roughly 500 to roughly 8,000 events per second without violating ordering, atomicity, or replay safety — and a diagnostic framework for evaluating future scaling changes against the same correctness criteria.