Selected work · Architecture notes
I build agent runtimes, local AI systems, voice interfaces, and the product layers that make them useful. I care about useful autonomy, explicit safety boundaries, observable behavior, and software that still makes sense at 2am.
| Layer | Current system | Focus |
|---|---|---|
| Runtime | Base | Profile-driven agents, guarded tools, durable runs, MCP, memory, and a Go control surface. |
| Product | Data Analyst Studio | Local-first analysis workspace with browser Python, sandboxed artifacts, charts, and agent-assisted SQL. |
| Interfaces | Voice + edge experiments | Real-time speech, physical interfaces, and small hardware that makes software feel present. |
local-first — keep capability close to the operator when possible.
inspectable by default — logs, state, and decisions should be understandable
without archaeological expeditions.
guardrails before bravado — agent capability needs explicit boundaries,
not vibes and a long apology thread.
| Project | What it explores |
|---|---|
| simple-mcp-servers | Small, composable MCP integrations. |
| xevoice | Voice-interface experiments. |
| esp32-transcriber-soln | Edge transcription on constrained hardware. |
| doc-to-rules | Turning human documentation into usable system rules. |
| aitools | Practical utilities for AI workflows. |
Building durable systems from runtime to interface. Fewer demos. More receipts.