Chris Schuermyer

My partner and I took the city’s seven sample questions and expanded them to 14 that a homeowner or small business would ask (tiny houses, fence heights, cesspools, sidewalk signs), each tied to a real Portland address. Claude answered each one three ways: on its own, with our MCP server over PortlandMaps and the zoning code, and with the server plus a better-formed question. Every answer was split into individual claims, 453 in all, and each claim was checked before scoring.

Tools made answers more complete and less well cited: completeness rose from 2.21 to 2.86 while citations fell from 2.00 to 1.79. A better-formed question brought citations to 2.93 and completeness to 3.00. For anyone adding tools to an assistant: score citations separately, because here they dropped while accuracy and completeness improved.

An AI surgical-robotics startup needed a working demonstration for the accelerator’s Demo Day on May 14, 2026. I was the sole engineer and codeveloped it with a surgeon: she set the requirements and reviewed the outputs, and I rebuilt the prototype between sessions.

The system replays recorded surgical video as if it were a live camera. Gemma 4 reasons about the scene, SAM 3 segments what it names, and a dashboard shows risk indicators such as possible bleeding and poor visibility. Both models run on a single NVIDIA DGX Spark on-premises, to match a hospital’s concern about operating video leaving the building. An earlier version used a different vision model and a Next.js dashboard.

It is a demonstration on recorded footage, with no clinical accuracy claim.

Lives Lived is my own product: photograph a letter, journal, or postcard and get a transcript, a narration, and a graph of the people and places it mentions. Transcription runs on Gemini in the cloud or on Gemma on the phone, which keeps the pages on the device.

I built it alone with AI coding agents: about 250 Swift files, about 2,600 unit tests, and six App Store releases in the four weeks after launch.

Before recommending anything, I built a benchmark from about 100–150 of the company’s past support conversations and had a separate model score each candidate reply against what their staff had sent.

The scoring exposed a problem outside the model: some answers depended on facts missing from the company’s knowledge base. Those gaps went back to the client, and every failure their staff reported became a new test case.

The assistant shipped to the support team as a browser extension. Newer staff found it helpful; experienced staff were sometimes faster typing answers they already knew.

I was the technical lead alongside Sales on 10 to 15 head-to-head evaluations and renewals, and we won most of them.

An account, under two weeks. A major customer was weighing the competitor ahead of an executive decision. On site I found three causes behind their poor results: steps missing from their workflow, data one of their teams was not sharing with another, and bugs in our software. I helped their engineers make the case for the data internally, ran fixes with our U.S. teams overnight, and showed the improvement before the deadline. The customer expanded the contract.

A dispute, two companies. A chip’s designer and its manufacturer could not agree on the cause of low yield on a critical product. Working under NDA with both, I used our diagnostics to show that each side contributed part of the cause, and what each had to change.

The other half of the job was applied ML research on semiconductor yield, which produced four issued patents.