04 / Datambit · AI and media integrity · 2024
Genui: a detection model, and everything it needed to become a product
Datambit had trained a model that spots manipulated video, audio and images, and nothing around it. In five months it became Genui: a backend that absorbs bursts of very large files, a waiting estimate that teaches itself, a web client to work from, and a production environment that rebuilds from a single command.
- Jul 2024 Engaged, and Genui designed around bursts of work
- Aug 2024 Uploads, reports and support tickets running on the cluster
- Oct 2024 Wait estimates that teach themselves from finished jobs
- Dec 2024 Front end rebuilt in three days, then handover
The problem
Datambit’s business is telling you whether a video, an image or a recording was generated or manipulated. They had the hard part: a trained model, benchmarked against a national challenge. What they did not have was a product. No way for a customer to hand over a file, no way to know when the answer would arrive, no report at the end, and nothing an enterprise buyer could log into. The platform that became all of that is Genui.
The work also arrives in the least convenient shape. Nothing for hours, then a burst of files the size of feature footage, all at once, from a customer who expects an answer today.
What we built
A platform that absorbs the bursts
Uploads are taken in one piece, however large, and handed to the detection models as work to be picked up rather than as a request someone is waiting on. Quiet hours cost quiet-hour money, and a sudden batch does not take the service down. Because running a model costs real money per file, each account has a fair allowance, counted per person rather than per office, so an enterprise customer behind one company connection is not throttled as though they were a single user.
A wait time that teaches itself
A job with no estimate feels broken, and a wrong estimate is worse than none. So the platform learns from itself. Every week it looks back over the jobs it has actually finished, works out what a file of that size and kind really takes, and quotes the next customer from that. As jobs complete, everyone still waiting sees their estimate shorten. Nobody tunes it, and it gets more accurate the more the platform is used.
An environment that comes back
The entire production setup, machines, database and networking, is written down as code. One command rebuilds it from nothing, which is the difference between a serious outage being a restore and being a rescue.
The hard part
Very large uploads were failing in a way that looked like a network fault and was not. The order of operations inside the submission path meant a connection had to survive for as long as the file took to arrive, and over a domestic line it did not. Reordering the work so that nothing fragile waits on something slow fixed it, and the upload path was given room to take as long as a genuinely large file needs, because a customer sending a feature-length video is making a reasonable request and should not be treated as a stuck one.
What it changed
In under five months Datambit went from a model on a benchmark to Genui, a live platform around it: upload a file, see when it will be done, get a report. The detection was theirs. Everything a detection model needs before it can be put in front of anyone, and the infrastructure to stand it up again from nothing, was designed and delivered as one piece of work.
Datambit has since taken Genui private, for their own use behind their own doors, and the public interface is gone.