Clipmine

"Your videos are sitting on a goldmine. Start digging."

Overview

A creator with 200+ published videos is sitting on thousands of hours of content that stopped earning attention a week after upload — while needing fresh short-form clips every day. The existing tools (Opus Clip, Vizard and friends) all work on one video at a time. Nobody searches the whole library.

Clipmine does: ingest the back catalogue, transcribe it, embed it sentence by sentence, then answer questions like "find me the best 60 seconds about AI across all 800 episodes" — and cut and export the result.

What got built

• Ingestion pipeline — yt-dlp pulls the catalogue, Whisper transcribes, sentences get embedded into pgvector, all through ARQ background workers
• Cross-library semantic search at sentence granularity, which is the actual wedge — per-video tools cannot answer a question that spans the archive
• Clip assembly and export with watermarking, storage on Cloudflare R2, share links
• A complete SaaS shell around it: Google OAuth, multi-tenant data model, Stripe billing tiers, referrals, lifecycle emails, PostHog and Sentry
• 11 API routers, 14 services, 14 frontend pages, ~6,800 lines of backend

The honest part

All of it was built between February 11 and February 14, 2026 — four days — and then I stopped.

I keep it on the wall because the interesting question isn't the product, it's the velocity: what does a single person plus agents actually produce in a long weekend, and where does that approach break? Two answers from this one:

• It produces a lot of surface — every screen, every tier, every email — very fast.
• It skips the things nobody asks for. There are no tests, and the last commit before I stopped is titled "Fix multi-tenant auth bypass, ARQ watermark bug, and 19 other findings" — a self-run audit that found a hole where one tenant could reach another's data. Built in four days, and it needed a fifth to be safe.

It never shipped. No users, no revenue, no deployment — a 1,700-line PRD and a working prototype that answers the question it set out to answer.

Why it's here

Because a portfolio of only the things that worked is a portfolio that lies about how building actually goes. This one is a genuinely good idea, executed fast, and parked deliberately — and the auth bug I found in my own code is more instructive than another feature list.