AI & automation
AI tool that turns saved videos into a searchable library
We built a tool that turns the short videos you save on your phone into a searchable, summarized library. It is a working prototype that we have not deployed yet.
A tool that captures the short videos you save, transcribes and analyzes each one with several AI models, and files them into a searchable, summarized library with a weekly digest. We built it in one day to test a pattern we use in client work.
- One-tap capture
- Transcription
- Visual analysis
- Search by meaning
- Weekly digest
- Next.js 16
- Supabase + pgvector
- Inngest
- Vercel Sandbox
- Groq Whisper
- Gemini
- Kimi K2
- Resend
Before
This one we built for ourselves, to prove out a pattern we bring to client work. Saved videos vanish into a feed nobody revisits. We built a process where one tap from the phone's share menu sends a video in, then it is transcribed, its frames are read, it is summarized, tagged, and scored, and it is filed into a library you can search by meaning, with a weekly summary by email.
What made it hard
Capture had to be one tap from the phone's share menu, with no app to open.
Each video had to be transcribed, watched, and understood, which means coordinating several AI models, each good at a different part of the job.
Search had to understand meaning, not just keywords, so a video can be found by what it was actually about.
What we built
Built a share-menu capture that hands each saved video to a background process, with a setup guide for the phone shortcut.
Coordinated four AI models end to end: speech-to-text for the audio, a vision model for the frames, a reasoning model to summarize, tag, and score, with backup providers for that step, and an embedding model for search.
Made the library searchable by meaning with vector search, so a video can be found by what it was about rather than its caption.
Grouped saved videos into topics and set up a weekly digest of the highlights, with an hourly retry for anything that failed.
Results
Speech-to-text, a vision model, a reasoning model with backup providers, and an embedding model for search.
Vector search finds a video by what it was actually about, not just the words in its caption.
Nine commits on a single day, May 28, 2026, including tests.
Where it stands
Prototype. Built and tested, not yet deployed.
