Two consulting engagements, on other people's platforms.

Personal-injury firms reconstruct a medical chronology by hand from thousands of pages of records before they can write a demand letter. The work built a distributed multi-store AI document platform that turns raw legal-claim documents into a structured settlement knowledge graph: OCR and page classification run first, then an LLM extraction graph with a LangGraph state machine covering 11 document categories, with spaCy for medical terms and ICD-10/CPT mapping, pgvector for retrieval and Neo4j for entity relationships. A custom token-bucket rate limiter enforces per-model RPM/TPM with exponential-backoff retries, so the extraction graph stays inside its throughput budget under load. The whole thing runs as cooperating Python services over RabbitMQ with per-tenant data isolation for legal compliance.

Job-market intelligence is a distributed data-engineering problem before it is an analytics one: the whole posting market has to be processed continuously, not sampled. The work built a job-market data platform on Ray, Kafka, Iceberg, ClickHouse and Redshift, the entire AWS footprint provisioned and versioned in Terraform on EKS. Kafka on Amazon MSK carries new postings from the crawler tier to Ray workers over an Apache Iceberg lake, Redshift with Spectrum serves the warehouse, FastAPI serves the dashboards, and a ClickHouse migration removed a data processing lag of roughly 3 weeks.

The five systems built solo carry the same method, in five unrelated industries, built end to end by one engineer. These two engagements ran inside someone else's existing platform team, on someone else's platform, under someone else's compliance regime. Both kinds of work are real and both are listed here, but they are not the same claim, and the fence is what stops a reader from mistaking one for the other. The systems table on the Work page is the first row; this page is the second.