Search and large language model infrastructure in Python and Go, across distributed microservices talking over Apache Thrift RPC.
Designed and shipped an autonomous agent that maps a live admin surface into a graph of pages and actions, ranks journeys through it against real user click analytics and emits executable seeds with no hand-written scripts. It took reachable-state coverage from 5.9% to 15.6%, and every fix it surfaced is pinned by a regression test. Adopted by 3+ teams after senior engineer review.
Hardened the agent after a live run acted on a production record, on a database shared with pre-production. The guard was rebuilt fail-closed and view-only by default: mutations on existing entities are blocked outright and the final create is gated behind a second explicit flag, each rule pinned by a test that runs offline without a live browser.
Built an LLM service that turns requirements documentation into executable cases, and authored 12 agent skills covering the full project lifecycle for the team's assistant bot. Demoed at standup, taken up by other teams, and now used to answer production data questions that used to take a manual investigation.
Diagnosed failures across a multi-region microservice deployment covered by an 870+ case Python suite, tracing them to ACL configuration, service reachability differing by data centre and interface definition drift rather than the assertions that surfaced them. Recovered 50+ blocked cases by tracking a broken delete endpoint to an Apache Thrift IDL mismatch between client and server schemas.
Built and deployed a health-monitoring service across 7 regions and both production and pre-production. It polls an operations API and pages on-call engineers through a chatbot, has caught 5+ multi-region incidents, and I closed a hole where a total outage could report zero targets down.