YouTube Discovery & Classification Pipeline
A multi-keyword search and filtering pipeline that finds quality tutorial content at scale, judges it with Gemini, and files it into a nested category tree automatically.
Multi
Keyword search matrix
2-stage
LLM quality gate
Nested
Auto-categorisation
Client
Client name withheld at their request.
How it works
- 1
Searches many keyword permutations against the YouTube Data API with tuned parameters for duration, recency, language and engagement.
- 2
Dynamic filter logic scores each result, discarding low-signal uploads before they ever reach a model.
- 3
Gemini evaluates the survivors for genuine instructional quality — not just view count — and rejects clickbait.
- 4
Newly discovered channels are scraped and passed to an LLM that decides whether the channel is worth tracking at all.
- 5
A second Gemini pass assigns each video to a nested category tree, so the library stays browsable as it grows.
- 6
Approved records are inserted directly into the web app's database.
Tech stack
Want to see it running?
I demonstrate the full workflow live on Google Meet — the canvas, the logic, and the system handling real input. Happy to answer anything the write-up didn't cover.
I don't share workflow JSON. These are custom builds and the architecture is the work.
