Led engineering for an access-control system on a customer site in Japan — an in-house vision API deciding who may pass through a secured door, and commercial lock hardware that had never heard of it. The model was the easy half. A vision model returns a confidence score and a lock needs a yes or no, and the yes is physical and irreversible. Where that threshold sits, and how it fails when the camera is dazzled or the network drops, is the actual system.

Ran daily in production, on a site I had no control over

Built a Claude Code workflow that assembles our engineering status reports from the tools that already hold the truth — the issue tracker and the repositories — attaching evidence to every line it writes. The design question was which judgments stayed with a human — the agent gathers and drafts, people decide.

Most of a day of manual collation down to well under an hour, every claim traceable

Delivered a companion robot prototype running a speech-to-text → LLM → text-to-speech loop. Latency is the whole product here: a pause that would be unremarkable in a chat window is a broken conversation when someone is speaking out loud to a machine.

A production database had outgrown the host it shared with other services and started starving them of memory. Moved it to dedicated infrastructure via replica-set promotion with a scripted cutover: quiesce writers, verify sync, promote secondary, rotate credentials, restart.

Ten seconds of unplanned downtime, inside a single planned window

An eight-year-old platform was built on the premise that the identifiers it received from upstream were stable. They stopped being stable, so records that were one thing started arriving as several. Rather than rebuild, I added a content hash over the attributes that actually identify a record as a fallback key, checked only when the primary ID misses — fixing integrity without breaking existing API contracts.

Integrity restored without breaking a single published contract

I owned a high-throughput fetching platform — hundreds of thousands of jobs a day across a few hundred workers — and its core was the bottleneck. Migrated the core scraper from Python to Go and rolled the rewrite into the live system.

10x execution-time improvement, shipped into production traffic