Projects
What I'm building in public.
autoplay
A local autonomous Stardew Valley agent: a C# SMAPI mod exposes bounded observations and controls over a named pipe, a two-model Python harness (actor + director) drives it through OpenRouter. Objectives complete only when structured game state proves it, never the model’s own account of what it did.
ai-usage-heatmap
A GitHub-style heatmap of your AI token usage, on your GitHub profile. One npx command turns a tokscale or ccusage export into a stats card with daily totals and agent icons, committed to your own repo. No server, no signup, zero dependencies; your usage data never leaves your machine.
neuromancer-mcp
Most hiring tooling makes a candidate's agent apply or a recruiter's agent screen. This flips it: my résumé is a remote MCP server a company's agent connects to with one command, to read my work, weigh it against a role, and reach out. Eight tools over Streamable HTTP, inbound-only contact, built to be forked by swapping a single file.
hermes-local-agent
A personal AI agent running entirely on a 2018 gaming laptop with 6GB of VRAM: Hermes + Ollama + qwen3.5, driven from Telegram. Five measured budget constraints decide a local agent stack, and the KV cache, not parameter count, is the real bottleneck. Blog-only build, every number measured on the laptop.
prompt-cache-economics
A timestamp at the top of your prompt is a 0% cache hit on OpenAI, and on Anthropic it costs more than not caching at all. This repo measures prompt-cache economics across providers, with a recorded benchmark and an interactive replay you can step through.
agent-coercion-layer
Swap models in a production agent and their output shapes differ: one fences its JSON, one truncates a tool call at a nullable field. This repo measures what differs across models, and ships a transparent coercion layer that absorbs it.