Applied systems
AI
Design and ship AI features, agent workflows, and model integrations that hold up in production: retrieval, evaluation, cost control, and the boring operational pieces most prototypes skip.
Discuss this workThe usual failure
Most AI work dies between a notebook and a reliable product surface. Latency, evals, prompt drift, data boundaries, and cost show up after the first demo — when it is already in front of users.
How I take it
I start from the job the model is supposed to do, then design the thinnest system that can do it safely. That usually means clear interfaces around providers, retrieval, tools, and human review — plus instrumentation so you can tell if it is actually working.
What you leave with
▸Feature-shaped AI, not a chatbot bolted onto the homepage
▸Retrieval, tool use, and eval loops you can operate
▸Cost, latency, and safety constraints designed in
▸Handoff docs a team can keep evolving
Typical engagements
▸AI feature architecture and implementation
▸RAG / agent workflow design
▸Provider selection and abstraction
▸Evaluation and production hardening
Other lanes
If the work is real, I will take it.
Pactum is a one-person practice. You work with Matthew Swezey — not a bench that gets swapped after the pitch.