AI
captr
Point it at a topic and it builds you a research doc — searching, transcribing, and frame-grabbing video into something you can read and then chat with.
Research from video is a slog — you scrub through hours to find the ninety seconds that mattered. captr does the scrubbing. Give it a topic and it goes out, watches the videos for you, and hands back a written, illustrated research document you can then interrogate like a chat.
What it does
- Topic in, queries out — Gemini turns your topic into a set of YouTube search queries worth pulling.
- Watches so you don’t — it downloads and transcribes the results (yt-dlp + transcript APIs), then uses the transcript to find the moments that matter.
- Grabs the visuals — FFmpeg extracts key frames at those highlight timestamps, so diagrams and demos make it into the doc, not just words.
- Assembles and answers — everything is stitched into a Markdown research doc backed by the source material, which you can then chat with for follow-ups.
The pipeline
The interesting engineering is orchestration — a multi-stage pipeline where each step (search → transcribe → locate highlights → extract frames → synthesize) feeds the next, with the LLM used for judgment at the ends and plain tooling doing the heavy lifting in the middle.
Behind it sits a normal, real app: JWT + bcrypt auth, transactional email via Resend, and MongoDB persistence with search — with global exception handling and a consistent response envelope so the frontend never has to guess.
What I’d do differently
Frame selection is timestamp-driven; a lightweight visual-dedup pass would stop near-identical slides from all making the cut and keep the docs tighter.
The goal was to turn “watch these ten videos” into “read this one page” — and keep the receipts so you can always jump back to the source.
Want something like this built?
I take on freelance backend and API work — auth, payments, media, and the security review to go with it.
Start a conversation →