SCM Finds Video Clips by Cue: Vectors for Scenes, OCR for Text, Whisper for Lines
sven_ai · x · 2026-10-14
SCM is a local video search tool for Mac built on the idea of picking the retrieval method per cue instead of throwing everything at an LLM:
- Visuals: split video into scenes and locate shots with visual embeddings
- Text: OCR handles on-screen text
- Dialogue: Whisper transcribes speech for literal matching
Results jump straight to the matching clip. Models run offline after download and media never leaves the machine.
Caveat: indexing samples by density rather than understanding every frame, so brief on-screen moments can be missed.
More from coding & agent
- PowerShell demo shows AI workflow to classify, rank and route requests — dfinke · 2026-10-15
- Expand Fetch launches: agents see the web, hits 74% vs Exa's 55% on hard extraction benchmark — bclavie · 2026-10-15
- Agent breaks prod, then fixes it itself: 502, Coolify logs, self-recovery — tristanbob · 2026-10-15
- With zero motion design background, dev crafts Apple-style launch film in 3 days using Opus 5.5 — Glittering-Art3723 · 2026-10-15
- iMCP 1.7.0 adds HomeKit control, voicemail transcripts, and editable Reminders for Mac MCP — pcuenq · 2026-10-15
- Snowflake FDE Podcast: Same Model, 25% to 85% Accuracy — Inside Frontier Engineering — ivory_tang · 2026-10-15