What it is
Flow is a personal feed that reads from a curated set of sources — YouTube channels and RSS — and ranks each item by how close it is to the projects and topics I'm working on, instead of by how many people clicked it.
The relevance is semantic, not keyword-matched: a topic map plus sentence-embedding similarity decides whether a video about "agentic workflows" belongs near my Kanade work or not. The feed is deliberately bounded — a short, ranked list I can actually finish, not an infinite scroll.
Why I built it
Every recommendation surface I use is optimized for watch-time, which means it quietly fights me whenever I'm trying to stay on one thing. I wanted the opposite: a feed that knows what I'm building and brings me the handful of videos and posts that actually move that work forward, then stops.
What I learned
- Discovery and ranking are different problems. A feed can order a pool perfectly and still be useless if the pool was never built to contain relevant items. Curating the sources mattered more than tuning the score.
- Topic matching has to understand meaning. Matching my private project names against video titles found nothing; mapping projects to topics and comparing embeddings found the right videos.
- Text-first beats fake audio. An early version narrated items with synthetic voice; reading a clean summary turned out to be calmer and more honest than pretending there was a podcast.
Status & next steps
Flow runs as a private MVP on the homelab, behind authentication — it's a me-only tool, not a public demo. The recommendation and discovery layers were recently rebuilt around a curated source list, a topic map, and embedding-based relevance. Next up is widening the curated sources and a longer evaluation of how well the ranking holds up over a normal week.