Useful AI is mostly constraint work.
I'm an applied AI engineer. I like the part after the impressive demo, when the system has to answer from the right source, refuse the wrong action, stay inside a budget, and be owned by somebody when it wakes up broken.
My work spans enterprise agents, retrieval over private knowledge, local-model experiments, self-hosted deploys, CLIs, and web interfaces. At SPS Commerce I helped build the "Max" agent and its retrieval foundations; outside work I build the same ideas in public, smaller, and with more of the wiring exposed.
The model is rarely the whole product. The product is the permissions, retrieval, fallbacks, logs, pricing, deployment, and taste wrapped around it.
Build
Ground answers in something inspectable: notes, docs, citations, traces, and clear fallback paths.
Optimize
Constrain what the system can do: permissions, quotas, model choice, cost ceilings, and boring failure modes.
Ship
Operate it somewhere real: Proxmox/systemd, Cloudflare Workers, Vercel, CI gates, packaged tools, and runbooks.