Applied AI · Media
A scalable microservice that watches video uploads, runs computer-vision and LLM models to detect objects and write contextual descriptions back into the MAM turning terabytes of unlogged footage into searchable, reusable assets.
Overview
The client runs an enterprise media operation with terabytes of video footage living across Media Asset Management (MAM) platforms like Iconik and CATDV. Most of that footage was effectively invisible searchable only by filename and whatever a human had manually logged, which for the bulk of the archive was nothing.
Post-production teams couldn't find the footage they already owned. Manual logging doesn't scale to terabytes of uploads, so a large share of the archive went unused. The system had to identify what was actually in each video and make it discoverable without adding work for editors, and without bottlenecking on a single synchronous pipeline.
Approach
01
A Python microservice, containerised with Docker and coordinated over RabbitMQ, consumes video uploads as they arrive and processes them asynchronously so ingest never blocks on inference and the system scales horizontally under load.
02
Custom computer-vision models (Roboflow) detect objects in the footage, and an LLM turns those detections into contextual, human-readable descriptions the kind of metadata an editor would actually search for.
03
Generated objects and descriptions are written straight back into the MAM via its API, so the tags live where the footage lives inside Iconik or CATDV with no separate tool to learn.
Outcome
Manual logging was eliminated. Terabytes of previously unsearchable footage became discoverable by what's actually on screen, letting post-production teams reuse assets they already had and measurably increasing asset utilisation across the archive.
Stack
Thirty minutes. We'll listen, and we'll tell you plainly whether this is a fit. If it isn't, we'll point you at someone it is.
No discovery deck. No sales call. One conversation.