ClipFinder
Dec 2025 – PresentAI-powered video clipping for creators. Solo build, end-to-end.
I owned every layer: Next.js + React frontend, Trigger.dev task orchestration, a Python yt-dlp worker fleet on DigitalOcean, an FFmpeg server-side export pipeline, two generations of Stripe billing, a public REST API, a remote MCP server, and email drips on Resend. No co-founder, no engineering hires.
What it does
Paste a YouTube, Twitch VOD, or Google Drive video. Gemini analyzes it and surfaces the moments worth clipping. You preview, trim, reframe to portrait, burn in captions, add a text hook, then publish to YouTube or TikTok directly. There's also a free AI clip-grader, a public REST API, an MCP server, and a Talk-to-Video chat over the transcript.
What I built
- Async job orchestration via Trigger.dev v4. Per-segment retries with idempotency keys, wait-token callbacks to the Python worker, structured error envelope across the whole pipeline.
- Server-side FFmpeg export. Face-aware portrait reframe, burned-in captions (ElevenLabs Scribe with speaker diarization), text hook overlay, end card. All rendered server-side, not in the browser.
- Two billing systems, and the judgement call between them. First Stripe metered billing with Credit Grants, cancel-on-failure refunds, and atomic spend caps. Then, once the data said usage-based pricing was the thing capping revenue, a full rebuild to flat subscriptions with tier resolution and admission control enforced atomically in Postgres.
- Multi-source ingestion. YouTube, Twitch VODs, Google Drive, direct upload. Pre-flight probes that throw structured errors before a job is admitted, so failed submits cost the user nothing.
- Public REST API and remote MCP server (Supabase OAuth). The product is consumable by other agents, not just humans. Zod schemas generate the OpenAPI spec automatically.
- 204 SQL migrations, RLS on every table, three test tiers (unit, integration, expensive), husky pre-push hooks. Built like a production system from week one.
What happened
From Dec 2025 to July 2026, ClipFinder processed 12,272 hours of video and generated 87,521 clips for 8,452 users. Users went on to post 820 of those clips to YouTube, where they collected 1.86M views and 43,570 likes. Top single video: 69,432 views. Sustained roughly 1,000 to 1,400 signups per month for eight straight months on organic alone.
Revenue took longer to move. Usage-based pricing stalled around $300 a month through May. I replaced it with flat memberships in June, and MRR is $1,025 today on $3,614 lifetime. Same product, same segment, same traffic.
“We'd given up on clipping after seeing how long it took to select, caption, and export ourselves. Since switching, we've reached over 20,000 new YouTube viewers in a month, and publishing clips is actually fun.”
What I learned
I got the diagnosis wrong the first time. For five months usage went up and revenue stayed flat, and I concluded that short-form creators structurally can't afford tools like this. That read was wrong, and it was wrong in the most expensive direction: it blamed the market for something I controlled. I was selling metered pricing to people who wanted to know their bill before they clicked submit.
So I killed metered billing and shipped flat memberships. Nothing else changed. MRR went from roughly $300 a month to $1,025 in seven weeks, and signups kept climbing to 1,400 a month. The ceiling I thought was in the market turned out to be on my pricing page.
There is still a real constraint here. This is a small market with a hard cap on what any one customer is worth, and I'd rather point the same effort at a bigger problem with more leverage. But the lesson I actually take away is about diagnosis: I had the data to find this months earlier and I read a pricing problem as a market problem instead. ClipFinder keeps running, the infrastructure takes care of itself, and I'm looking for what's next.
