TECHNOLOGY
The bottleneck was never model capability.
It was false-alarm suppression. That single realisation is why Atlas was rebuilt from the ground up, and it determines almost every engineering decision in the system.
THE ORIGINAL APPROACH
Scoring every frame does not work.
The obvious way to build this is to sample video on a fixed interval, send each window to a large vision-language model, and ask it how dangerous the scene is. It is the first thing we built, and it fails in four directions at once: the false alarm rate is unusable, cloud latency puts the answer behind the event, inference cost scales with camera count rather than with incidents, and sending clinical footage to a third-party model is a non-starter before anyone even evaluates accuracy.
The instinct is to reach for a larger model. That is the wrong move. A larger model asked the wrong question more expensively is still asked the wrong question.
THE CASCADE
Separate finding candidates from judging them.
Perception runs continuously because it is cheap. Reasoning runs rarely because it is not. Each stage exists to protect the next one from work it should never have been given.
- STAGE / 01
Ingest
Every frame from every connected stream, decoded on-site.
- STAGE / 02
Perception
Detection and tracking
Objects detected and given persistent identity across frames. Cheap enough to run continuously on every stream.
- STAGE / 03
Trigger
Motion / spatial rules
Tracks tested against the geometry and timing of the environment. Only what looks like it might matter proceeds.
- STAGE / 04
Reasoning
Local vision-language model
A vision-language model reads the candidate in context and decides whether it is genuinely meaningful. Expensive, which is why it runs here and not before.
- STAGE / 05
Event
A structured event with a cause, a location, and a time. This is the only stage a person ever sees.
BAR WIDTHS ILLUSTRATE THE SHAPE OF THE CASCADE, NOT MEASURED RESULTS.
COMPONENTS
What runs, and where.
Perception
Object detection and tracking
Detection gives us objects; tracking gives them persistent identity across frames. Identity is what makes an event describable. You cannot say two people converged and stopped moving without knowing they are the same two people over time.
Trigger
Motion and spatial logic
Tracks are evaluated against the geometry of the environment and its expected timings. This stage is deterministic, inspectable, and configured per site. It is also where the overwhelming majority of activity is correctly discarded.
Reasoning
Vision-language model, on-site
Only surviving candidates are read in context by a vision-language model, on-site. Because it runs on a fraction of a percent of activity, a local model on modest hardware is sufficient, which is precisely what makes on-premise deployment possible.
LIMITS
What we are not claiming.
Every architecture has a shape, and every shape has a weakness. Ours are worth stating plainly, because you will find them anyway.
- What this architecture does not solve
- It does not make a vision-language model correct. It reduces how often one has to be asked, which is a different and more tractable problem. A candidate that reaches the reasoning stage can still be judged wrong.
- Where it is weakest
- Environments with no stable baseline. If normal genuinely changes every day, deviation stops being a useful signal and the system's advantage collapses toward that of a conventional detector.
- What we have not proven yet
- Long-horizon performance across many sites. The architecture is sound and the pilot is running, but we do not yet have the deployment count to claim that the reusability argument holds in practice. We would rather say so.
Want the detail behind any of this?
We are happy to go deeper with technical teams and investors who want to interrogate the architecture properly.