You have the video footage. Finding it is the problem.

A hundred cameras record without fail. The trouble starts the day something happens — a crate goes missing from the warehouse, someone badges into a room they had no business in, a client swears they were on site Tuesday. Now security has terabytes of video and one tool: a person scrubbing through it. Across a hundred cameras and a week of recordings, that person loses. They check the two cameras they have a hunch about and stop.

Digercules works the other end of the problem. We don't do live monitoring and we don't replace your cameras. We go through what they already recorded — after the fact, on an ordinary machine in your office. You don't get "video someone reviewed." You get a short list: the file, the timestamp, and the reason it's on the list. Your team watches forty clips instead of four hundred hours.

What we find

Things that changed, not just faces. Security rarely asks "find this face." It asks "there was a crate here, now there isn't — when did it go, and with whom." We split each recording into scenes and flag the moment a camera's picture stopped matching what it showed before. A missing object. A door that moved. A cabinet left open. A van that wasn't there yesterday. None of that needs face recognition.

Your people, and everyone else. Faces from the selected frames go into a gallery. Confirm an employee once and the system tells known from unknown across the whole archive — and across every job after this one, because the gallery carries forward instead of starting over with each export.

It's built for people who change. Each person is stored as a set of reference shots from different times and conditions, not one averaged portrait. Averaging is what breaks recognition: blend a face across fifteen years and it matches nobody. We don't need a clean frame either. We take the sharpest shot in each scene and decide across all of a person's frames, not one blurred one.

And only a human grows the gallery. A name goes on a face after your employee confirms it — the product has no "the system decided it was him" path at all. That's how it's built, not a setting you can flip. So far 430 faces have been confirmed this way.

A retrieval plan. Finding it is half the job. The other half is walking out with a copy that holds up. You get a written plan: which files, which intervals, where they physically live, what order to pull them in, and a checksum for every copy. Months later you can still show that the file you handed the police is the file that was on the recorder.

How it runs

Your video footage stays on your side. Either the whole job runs on your machine — inventory, scene analysis and face search are built for a plain CPU, no graphics card — or the heavy stage goes to one specific outside machine over a closed peer-to-peer link. Processing runs either entirely on the client's own hardware, or — if that hardware isn't powerful enough — is delegated to a specific external machine over a closed peer-to-peer channel (not a public cloud): the compute is physically located in the Caucasus and Eastern Europe, the client is always told explicitly which machine and which jurisdiction is doing the processing, and only text/structured results come back — never the source files. What comes back is tables, descriptions and sample frames — never your recordings.

Questions we get

"A hundred cameras, a month of recordings. Realistically?" Cost doesn't follow file size. We read metadata across the whole archive first, without decoding any video, then pull key frames only, and everything downstream works from those. An hour of recording takes minutes on one CPU. For a real number on your volume, run the pilot on one day of video — that beats any figure off a price list.

"That's employee biometrics. Legal will sit on it for six months." Two separate things. Scene-change search never touches a face and processes no biometrics at all — if the question is when the crate vanished, that module stays off. If you do want people identified, the gallery lives on your side with the recordings and never reaches us in any form, and every name in it was typed by your employee, not produced by a model. The legal basis is yours to set; we fit the work to it.

"Our recorder is non-standard. Odd file format." Expected. That's configuration, not a rewrite. This system was built to crawl other people's messy archives — your recorder is one more of those. What we need up front: an export covering any one period, and a note on how your files and cameras are named.

What we don't do

No live monitoring, no direct connection to your cameras. We don't replace your security system and we don't write legal opinions. We don't put a name on a face without your employee's say-so. These are choices, not gaps: we work the archive and hand the material to the person who decides.

Where to start

Pick one real incident. Send us the recordings for that stretch and tell us what you were looking for. You get back the episodes with timestamps and a retrieval plan. Compare it to what your own people found by hand — or missed. That's the honest test.