Things I have designed, built and released. Each card links to the source and, where there is one, to a write-up on how the project works and what was hard about it.
A self-hosted broker that leases models to every opencode session across cost tiers and providers, steering spend by subscription budget and failing over when a provider breaks.
Read the write-up on opencode-brokerSource code for opencode-broker on GitHub
Permission modes and a safety floor for an AI coding agent’s shell: static rules and credential guards settle most commands, a classifier that fails closed judges the rest, and secrets are redacted from tool output.
Read the write-up on opencode-guardSource code for opencode-guard on GitHub
A ledger of real incidents caused by AI-generated work, filed by an unattended agent that reads coding-session transcripts and records what broke, what it cost and the lesson.
Read the write-up on ai-incidentsSource code for ai-incidents on GitHub
Turns one multi-GPU machine into several gaming stations: Windows or SteamOS VMs with a real GPU passed through, on a self-healing Fedora bootc host run from a web console.
Read the write-up on TendrilPrivate repository
Picks the most varied, highest-quality photos of each person from an Immich library with ArcFace embeddings and clustering, and delivers them as training data for Frigate face recognition.
Extends TrueNAS SCALE’s TrueCloud Backup to Backblaze B2 and any S3-compatible provider through a fail-safe overlay patch that survives reboots and OS updates.
Read the write-up on truenas-truecloud-patchSource code for truenas-truecloud-patch on GitHub
The control plane for my AI coding agents across several machines: config sync, a cross-host session dashboard, and scheduled unattended jobs that each run inside a declared permission envelope.
Read the write-up on claude-fleetPrivate repository