Riel St. Amand · Senior software engineer
Production ML, full-stack product, native tools, and research across healthcare labs, Medicaid, a Fortune 100 support bot, a tire factory floor, Google Cloud, SAP data, and the NFL. The chart below has all 25 projects, plotted by how deep each one went.
25 projects · 5 industries · 7 disciplines
Each column is a project, drilled through the layers of the stack it touched. Hover a column for the project, or a discipline to see everything that went through it.
Browse all 25 as a list →Tools & AI4
Sports5
Healthcare4
Enterprise11
Manufacturing1
Depth
Tools & AIactive
A native video editor in Rust, for Windows and Linux, whose chat panel drives the agent CLI you already pay for. The agent calls the same operations you do, so every edit it makes is undoable — four crates behind two trait boundaries, FFmpeg and wgpu underneath.
Ctrl+Zundoes the agent’s edits, same as yours
Healthcareproduction
Five small networks, one per PCR target, gated by the same rules a clinical lab uses — and all of it running in the browser on a lab workstation, reading a 384-well plate in under a second. It had zero reported errors, while human technicians corrected about 4% of their own calls.
1M+production diagnoses, zero reported errors
Sportspaper
A forecasting benchmark where the answer is known by arithmetic. At the clock-killing first down, win probability says 0.97 and the clock says 1.00. Thirteen methods, foundation models included, and none that learns from average loss closes the gap.
0.03the residual no average-loss learner recovers
Sportsactive
A Discord bot that grew into four live sites for people who treat the NFL Draft as a year-round craft. There are no ads and no gambling. One TypeScript monorepo on Firebase and Cloud Run runs the draft platform, fantasy, a stats workbench, and the bot.
4 sitesone TypeScript monorepo, zero ads
About
I’ve spent six years shipping production systems across stacks: React and TypeScript products, Python services, .NET on a factory floor, production ML with real stakes, and lately a native Rust video editor. I land in an unfamiliar codebase, get productive quickly, ship, and then write down what happened so the next person doesn’t have to rediscover it.
I teach too: a four-hour ML workshop at KCDC, a six-part series at Core BTS, and talks at HIMSS and Dayton AI Day. In open source, I fixed a GRPO checkpoint-resume bug in NVIDIA NeMo RL (opens in a new tab) that silently zeroed KL regularization, and I maintain nflverse-ts (opens in a new tab), the TypeScript port of the nflverse.
Currently building Kinewright, and looking for senior roles where range matters more than a single job title. Here’s what I’m after.