Builds
A short shelf of things I've built to answer questions I couldn't leave alone. Start with TinyDesk for workflow design, human review, and evaluation. For workplace agents and employee coaching, see AI in practice →.
A small help desk for working a queue of fictional Ashmere Press tickets. TinyTalk Assist summarizes a ticket, suggests a category, flags what’s still missing, and drafts a reply you can edit. Nothing saves until you click save.
JavaScript front end, Python backend. Local Ollama by default, cloud xAI Grok if you opt in. In a small 12-ticket eval scored for factual errors, repeated questions, and unsupported promises, low-reasoning Grok gave 10 usable drafts and 2 that needed edits.
Screenshot: real interface, fictional ticket. The assistant panel replays a recorded Grok result, not a live call.
A small local chatbot that became a practical test of memory reliability. Explicit facts, fact history, a current profile, and source status stay separate from recent conversation so “what is current,” “what used to be true,” and “what was supplied” do not collapse into one transcript.
A portable behavior layer for AI coding agents. Clamp scope, challenge assumptions, prove the work with evidence, then clean up the residue — across Codex, Cursor, and Antigravity.
Creative technology
Turn a year of GitHub contributions into a 3D city. Every day becomes a city lot, activity becomes height, and the original contribution history stays underneath the spectacle.
Local AI · Privacy
A localhost AI console with Ollama chat, SQLite threads, saved facts, and local jobs. Research is an explicit action, with local-note and web modes and source records attached to each run.
A terminal World of Warcraft companion with tools for Blizzard’s official structured data. Local or cloud models share the same lookup loop, with tests for routing, tool-call history, and API failures.