I build and run AI agent infrastructure — a fleet of coding and ops agents that's been in daily production for over a year, across local GPUs and cloud, for my own work and for other businesses.
The repos below are real tools carved out of that work and rebuilt to stand on their own.
🖥️ spot-gpu-autoscaler — rent cheap GPUs across providers, autoscale on a pending-job signal, and enforce a hard $/hr cap with an out-of-band reaper that kills orphaned instances even when the app is down.
🔒 sandbox-render — render untrusted or LLM-authored HTML safely, and prove the containment holds (with a real escape-test harness) instead of assuming it.
📟 claude-transcript-feed — a live, flicker-free web view of a Claude Code agent, streamed append-only from its JSONL transcript.
🔁 anthropic-sdk-helpers — retry logic that tells a transient throttle apart from a real usage cap (retrying the wrong one is the cardinal bug), plus prompt-cache helpers.
Also here: a zero-dep Minecraft status board on a raw Server List Ping client, constrained-lab-lut for color-matching video via 3D LUTs, and throttled-worklane for rate-limited work scheduling.