I build AI products at the layers where they have to work. Retrieval, inference, and the evaluation that decides what ships.

Document AI for aviation and manufacturing technicians, running on the tablet in their hands. I own the product line and wrote the evaluation strategy behind it: 90 golden questions across three manuals, four retrieval metrics, and a harness engineering adopted as the regression gate for every change. For the Gemma 4 swap I ran the bake-off against Qwen 3.5 and ColGemma and wrote the decision rule that settled it. 85% citation accuracy, 1.86s to first token.

Most of my week happens in Slack, and the agents were already doing the drafting. So the artifact should get made where the team is, not written alone and circulated for comment afterward. Clippy drafts specs and code in a public channel, so teammates shape the work while it is forming instead of reviewing a finished blob. It spread company‑wide; our CEO turned out to be the heaviest user, and the product team runs on it daily. I built it on an open‑source agent base and own the durable execution layer that keeps a shared thread trustworthy. A duplicated answer costs more trust than a slow one, so that is where most of the ~400 commits went.

ML‑driven power allocation across Microsoft’s $4.5B Energy portfolio. The model found capacity we already owned but could not see, and row‑level monitoring caught rack clusters drawing over their allocation before they turned into outages. 1 MW of stranded capacity recovered in the pilot, $20M in new customer revenue within three months, 30% fewer outages, and 99.999% availability held throughout.

Consolidated feature engineering across Meta’s Integrity org: 30 enforcement teams, shared ML representations for 5B+ entities, 200+ models. Engineers chose to build on it, hitting 85% adoption against a 50% target, and once it crossed that line I sunset the legacy platform. Model development time down 25%, $60M saved. Most of the value came from removing infrastructure, not adding it.