ACADEMUN

A takeover begins below the features. First the servers were recovered, then made reproducible: every service into a Docker container, Nginx in front as reverse proxy and static edge, configuration in the repository instead of in someone's SSH history. The warp was already threaded — badly — so we re-dressed it in place, without stopping the classes running on it.

Onto that warp the fixes went in weekly passes: the FFmpeg pipeline rebuilt around a real job queue — one worker per encode, HLS ladders instead of one monolithic MP4, presets tuned to the source — resumable uploads that survive bad connections, segment caching at the Nginx layer so a popular lecture stops punishing the transcoder, and the LMS itself: courses, progress, the parts students actually see.

Inherited code gets beaten twice. Review concentrated on the paths that had already failed in production: encodes killed mid-segment, corrupt source files, disks filling during a transcode, deploys landing during a live class. Jobs became idempotent — a re-run overwrites cleanly instead of doubling — retries are bounded, and a failed encode lands in a queue a human reads, not a log nobody does.

ACADEMUN runs stable in production today. Lectures upload, transcode and stream without an engineer watching, encode failures take the bounded-retry path instead of a late-night phone call, and the pipeline that used to be the emergency is now the least interesting part of the system. The teams that walked away weren't wrong about the difficulty — only about whether it could be held. Cloth other hands abandoned, worn daily.