Garrett Macarthur

Open source self-hosted media server in Rust. One binary on hardware you already own, your movies and shows on every screen in the house. No cloud accounts, no ads, nothing to sell you.

Self-hosted software has an installation problem, and it is not the people who read documentation. A container runtime, a compose file, a database to provision and a reverse proxy narrows the population who can run it to the technical, and each of them hits a different failure on a different distribution.

One static binary, no runtime dependencies, no account anywhere, and no mechanism by which I could later charge for a feature you already have. Not a promise not to: no mechanism. Open source for the same reason, so the guarantee survives me.

Remote access becomes the user's network problem, and with no analytics the roadmap runs on whatever people bother to report. in development · dog-fooded with developers and families · any server hardware or Docker · web, Android/Android TV, iOS/tvOS, Tizen/webOS

Multi-tenant school CRM handling student enrolment and downstream account provisioning, with sync integrations to external providers. Education sector, supporting a very large user base.

A six-day silent failure. An upstream provider changed their response format without notice; my endpoint accepted the unrecognised payload, found no enrolments, and returned success. Uptime monitoring stayed green the whole time, and the affected schools were raising it with the other provider rather than with me.

A sync job that has stopped must not look like a sync job with nothing to sync. Schema validation at every external boundary, unrecognised responses fail loudly, and alerting on the absence of expected work rather than only on errors. Tests fail if the instrumentation is removed.

Six days of undetected downtime and the credibility of every green dashboard until the boundary work landed. 6 days silent · 0 alerts fired

The agent delivery pipeline behind Formist's client and internal work, named for the weave. Frozen-plan governance, orchestrator, implementer and verifier tiers, deterministic test and lint gates ahead of any model review, PR-based asynchronous approval.

The numbers I had been optimising against were wrong. Human wait time was being logged as model latency, so every tuning decision drawn from the old traces pointed in the wrong direction.

Instrument first, then rebuild. Plans freeze before implementation, deterministic gates run ahead of any model review, and approval happens asynchronously in a pull request rather than blocking a run.

A full rebuild of the pipeline. 96% of spend traced to context size, not model choice · 3 of every 5.5 hours was process, not work