Every Creator Is Sitting on an Unlabeled Archive. Running Gemma 4 Locally on a 5-Year-Old MacBook Proves a Service Business: Build the Index That Makes Their Footage Queryable in English.

Every creator has the same problem: footage they'll never edit.

The archive keeps growing. The time to edit it doesn't.

A SimbaStack engineer documented exactly this in a blog post published today. He runs a safari lodge in the Maasai Mara, shoots constantly with iPhone, DJI Pocket, drone, Nikon Z8, and Ray-Ban Metas, and has years of raw footage sitting on SSDs with names like "Mara june 2024 backup final FINAL."

The footage is not lost. But it is invisible — because "every AI video editor on the market assumes your footage is already labeled."

His solution: build the index first.

What he actually built

The tool runs Gemma 4 31B Q4 locally in LM Studio on a 2021 M1 Max MacBook Pro with 64GB RAM. Activity Monitor peaked at 50.89 GB of swap during the bulk run. The machine ran hot, the fans spun up, and it kept producing sidecars.

The per-clip pipeline:

• ffprobe extracts file metadata
• exiftool pulls GPS lat/lon/altitude from iPhone, DJI Pocket, and drone footage
• Nominatim reverse-geocodes the coordinates (free, no API key)
• ffmpeg extracts five evenly-spaced frames at 1920px
• WhisperX transcribes with word-level alignment and speaker diarization in 97 languages
• A vision model (Gemma 4 via LM Studio) reads the frames plus transcript and returns YAML frontmatter and a prose description
• The sidecar .description.md is written to disk next to each clip

The output: every clip gets a sidecar containing lighting enum, time-of-day enum, color palette, face embeddings, GPS coordinates, shot type, a prose description, and suggested use cases. The archive goes from IMG_1103.MOV to a searchable, English-queryable record.

"The whole thing is a Claude Code skill, about 1,400 lines of Python. Claude Code wrote almost all of it."

The cost comparison

The initial research pointed at a SaaS stack — Eddie AI for editing, Higgsfield for B-roll, Submagic for captions, Buffer for scheduling — that came to about $140 a month.

The local build came to $22.

The $22 covers DaVinci Resolve Studio and ElevenLabs for voiceover on clips that earned it. The indexer itself runs free, locally. The footage stays on-device.

The business hiding in this

The engineer solved this for himself. But the problem is universal.

Every photographer, videographer, travel brand, hospitality property, documentary crew, and content team has a version of this: years of footage, no time to edit, no way to search it.

Most of them have never heard of LM Studio. They don't know Gemma 4 can run locally. They can't write 1,400 lines of Python.

That is the service opportunity.

You build the pipeline once. You run it for clients. You charge for the index — and then for the editing workflow on top.

What makes this defensible

The local-first approach is the selling point, not just the implementation detail.

Travel brands and hospitality properties are protective of their footage. A safari lodge does not want to upload thousands of clips to a third-party cloud service. The same is true for any brand with location-sensitive or person-identifiable footage.

Running Gemma 4 locally means the footage never leaves the machine. The sidecars are plain text. They travel with the data when files move between drives. They survive if the indexer breaks tomorrow.

"The leverage is upstream. Build the index first, make the archive queryable in English, and the editor on top becomes a thin layer doing what it was designed to do."

That sentence is the pitch.

How to start

You don't need to solve everything at once. Start with one client, one SSD, and the pipeline.

Offer to take their footage folder and return a searchable archive. Charge for the time and the compute. Keep the sidecar format simple: YAML frontmatter plus a prose description, one file per clip.

Once the client can search their footage in plain English, the next sale — editing workflow, Resolve integration, social publishing schedule — is obvious.

The index is the wedge. The workflow is the retainer.

Sources:

https://blog.simbastack.com/indexed-a-year-of-video-locally/

https://news.ycombinator.com/item?id=48222733