Indexing a year of video locally on a 2021 MacBook with Gemma4-31B (50GB swap)
3 hours ago
- #AI video indexing
- #digital archiving
- #local LLMs
- A developer splits time between running a lodge in Maasai Mara and intensive coding work, facing a backlog of unedited travel footage.
- Instead of using expensive AI video editing SaaS, they built a local-first indexer to make unlabeled video archives searchable via natural language.
- The system uses local models (Gemma 4 31B) to analyze frames, extract metadata, and create sidecar files, enabling querying for specific clips.
- Running on a 5-year-old MacBook, the process pushed hardware limits with heavy swap usage but proved feasible for bulk indexing.
- Key lessons: enums prevent model confabulation, local models close most gaps to cloud AI, and the real bottleneck is indexing, not editing.