Foma — Deployment Case Study

A working studio's 270 TB archive, answerable in plain language

A real single-tenant deployment of Foma inside a motion & post-production studio — what it connects, how it runs in production, and where the studio's data does and does not go.

Updated 16 August 2026 · product version 1.1.0 · in daily studio use since May 2026

At a glance


Indexed	270 TB · 2.11 million files across four storage roots
Storage	On-premise NAS, an on-premise 125 TB production server, and a cloud filespace
Systems unified	Media storage · project tracker · review platform · cloud documents · team chat
Ways to search	By name or path · by meaning · by what is said · by what is in the frame · by reference image
Running in production	24 background services · nightly and hourly indexing · a correctness suite that runs every morning
Privacy	Embeddings, thumbnails and transcription all computed on-site. Media never leaves the building.

The environment before

A working creative studio leaves an enormous, messy trail of data. In this one:

270 TB and over two million files — footage, project files, exports, stills — spread across a local NAS, a second production server and a cloud filespace.
Critical context lived in five systems that did not talk to each other: the file servers, the cloud drive, the task manager, the review platform, and shared documents.
Answering "where is the final 4×5 export of this campaign" or "which projects are quietly eating the active server" meant opening five applications and depending on whoever happened to remember.

There was no asset-management layer tying it together. Rushes, proxies, deliverables and review notes sat in separate systems, invisible to the people doing transfer, transcode, delivery and archiving every day.

The knowledge was all there. It just wasn't askable.

What was deployed

Find

By name or path — exact lookups across two million files, instantly.
By meaning — semantic search over filenames, folder structure, documents and transcripts, rather than keyword matching.
By what is in the frame — five frames are sampled from every supported video and understood by a vision model running on the studio's own hardware, so a shot can be found by its content when nothing in the filename describes it. A reference still, screenshot or moodboard frame can be dropped in to find visually similar footage.