A production product, a working healthcare prototype, and an agentic operations system in development. Each shows how I move from a real workflow to an implementable architecture, make practical technology choices, and state maturity and tradeoffs plainly.
Selected work
Commercial proof, technical depth, honest status.
MakeCalls is the commercial proof point. Kintsugi demonstrates a multi-model healthcare workflow. Sweetwater Ranch shows how I structure an agentic operating problem before calling it production.
Commercial voice AI
Production product
MakeCalls
AI phone reference checks for landlords
An AI voice agent phones a rental applicant's prior landlords, holds a structured reference conversation, and Claude turns the call into a written report. A live product with real calls and real payments.
What it provesA narrow business workflow can become a complete AI product, including voice infrastructure, structured reasoning, reporting, and payments.
Three independent AI systems read one voice sample. An acoustic model scores how it sounds, transcription captures what was said, and Claude reconciles the two, naming where they agree and where they diverge.
What it provesIndependent model outputs can be separated, compared, and made legible instead of collapsing uncertainty into one confident answer.
Read the build notesOpen the live demoPassword-protected; access on request. Experimental only: not HIPAA-compliant, diagnostic, or used with real patient data.
Agentic AI
Prototype in development
Sweetwater Ranch
Agentic operations triage on a simulated ranch
A prototype in development designed to watch about 160 synthetic sensors across a fictional 34,000-acre cattle ranch: water, feed, fence, fuel, wells. In the design, code filters every sweep for free, the model reasons only about newly opened incidents, and each one becomes a work order. Four independent APIs expose 19 MCP tools to the agent. It is not deployed, publicly available, or production-ready.
What it provesAgentic architecture starts with boundaries: deterministic filtering, deliberate model calls, explicit tools, and an operational unit of work.