Kevin Rich
Building ai-driven, cloud-native applications — and the high-performing engineering teams that ship them.
A self-hosted personal video media server that reads each file's own embedded metadata as the source of truth, built in Go with a SvelteKit frontend.
Most media servers make you bend your library to fit them — rename files to a scheme, maintain a database by hand, or hand your collection to an online service to scrape. Holodex takes the opposite stance: the metadata already embedded in your video files — titles, cast, genres, dates, written into the container’s iTunes atoms or Matroska tags — is the catalog. Point it at a folder, and it indexes everything with no naming conventions, no internet connection, and no telemetry.
Media is a long-running hobby of mine and my university minor, and I’ve built versions of this project for years. This time I paired with Claude Code and reached a working MVP in a few days — work that has taken me a few weeks in past attempts. I deliberately chose Go because I’d never used it before; in practice my approach left me little need to drop into the Go code by hand.
The backend is a single pure-Go binary: a layered exiftool + ffprobe extraction pipeline, SQLite with FTS5 full-text search (diacritic-folded, so amelie finds Amélie), and a chi REST API. Indexing is incremental — the scanner watches the filesystem and only re-reads files whose size or mtime changed, so rescanning an unchanged 10k-file library makes zero subprocess calls. Cover art is tiered: embedded poster art is pulled instantly at index time, while files without it get a throttled background frame thumbnail so a modest host never gets hammered. The frontend is a SvelteKit SPA on semantic design tokens that drive three fully switchable skins from a single header control, with no per-component restyling.
It’s built to be run, not hosted-for-you: the library mounts read-only, the index and config live in a named volume, and the whole thing ships as one multi-arch image (amd64 + arm64) for NAS and ARM home servers. Past the MVP it now runs an MCP server that hands the library to AI assistants, a “System Activity” panel for the indexing internals, and a first pass at enriching cast from external metadata providers into a provenance-tracked shadow layer — editing and writing that metadata back to the source files is what’s next. Architecture decisions are tracked as a numbered ADR series in the repo.