At Harvey AI, I moved external side effects out of database transaction blocks and eliminated a class of production transaction-safety violations, from ~900 a day to zero. I also used query-plan analysis to cut an endpoint's latency from 1.5 seconds to 50 milliseconds.
I built Harvey AI's context-selection feature end to end, spanning Flask and SQLAlchemy APIs, OpenAPI-generated contracts, and a React and TypeScript frontend. I also built a shadow-mode evaluation harness for LLM output and safety tests covering cross-tenant data leakage and prompt injection.
As a freelance Full-Stack Engineer, I shipped nine products solo for paying clients, from scoping through production; five clients returned for maintenance. That work included subscription billing and prepaid API credits with idempotent payment webhooks, reconciliation, and revenue reporting.
I also developed a production logistics SaaS that collects cargo orders from websites and Telegram channels, processes them through queued pipelines, and extracts document data with OCR and LLMs. Its document extraction reached ~90% field-level accuracy on ~200 documents per month.
