Developer Indexes One Year of Video Locally Using Gemma 4 on a 2021 MacBook Pro
Tags AI · Consumer · OSS
A developer built a local video indexing pipeline using Gemma 4 31B on a 2021 MacBook Pro M1 Max (using 50GB swap), processing one year of unlabeled video footage from multiple cameras (iPhone, DJI Pocket, drone, Nikon Z8, Ray-Ban Metas). The pipeline combines ffprobe metadata, exiftool GPS, Nominatim reverse-geocoding, ffmpeg frame extraction, WhisperX transcription, insightface face detection, and vision model description to generate searchable .description.md sidecar files. Built using Claude Code with Opus 4.5/4.6.
Technical significance
Running a 31B parameter model locally on consumer hardware (M1 Max with swap) demonstrates how quickly local AI inference is improving. For professionals managing large media archives — journalists, researchers, content creators — local indexing eliminates cloud upload costs and privacy concerns while enabling semantic search over personal video libraries.