Field note · 02

What runs today, and what is designed

The maturity of each capability we describe, stated plainly. What runs in our recognition workspace now, what is designed for the edge and the airframe, and the rules we hold to before publishing a number.

Marketing sites for recognition software tend to blur three things: what runs, what is designed, and what is hoped for. We keep an internal register that ties every capability statement on this site to its basis and its maturity. This note is the public form of that register.

The rules we publish under

No accuracy, latency or throughput figures. Internal model reports exist, but on narrow class sets and internal validation sets. None is published until there is a defined evaluation, with a dataset, a sample size, a configuration and a scope, that holds for the deployed configuration.

Ownership stated precisely. The full-frame detectors in our workspace are independent, established detector families. The verification layer, the consensus and association logic, the evidence model and the integrity controls are our own work, and the site says so in those words.

“Running today” means running today. It covers uploaded footage and stills, multi-model detection, verification, temporal association, visual search with geometric verification, evidence and audit, model integrity and operations records. Live camera ingestion, automatic alerting and edge deployment are not claimed as running; edge is described as “by design”.

No customers, sites, people, infrastructure identifiers or hosting regions appear on the site.

Running today

  • Independent detectors run alone or in tandem with one-to-one overlap matching; raw observations are retained; context classes never veto. Running in pilot.
  • Candidates are cropped at original resolution and re-examined by a separately trained verification layer; scores are kept separate; there is no automatic all-clear. Running, labelled experimental.
  • Video is sampled on true timestamps, boxes are mapped to native pixels, and replay, findings and structured export exist. Running in pilot.
  • Same-class detections are associated across frames and the association resets on scene change. Running.
  • Visual search over private collections uses self-supervised embeddings with keypoint and homography verification, and the reviewer’s decision is recorded. Running; acceptance test passed.
  • The threat vocabulary covers firearms, edged and blunt weapons, explosives, and fire and smoke, with person and tool context. Running on an independent detector.
  • Hashes are declared at upload and verified before analysis; observations are immutable; human records are kept separately; access is private and signed; the audit trail is append-only. Running.
  • Models load only after an exact hash match against an approved manifest, and every result carries provenance. Running.
  • Sites, areas and camera inventory, an incident workflow with immutable history, role-based administration and single-use intake links. Running as records; notification delivery is locked off.

Designed, not yet fielded

  • The model-engineering practice (viewpoint-specific data, annotation rules, hard negatives, frozen validation, class-level metrics, gated release) is defined and tooled. Camera-domain fine-tuning has not yet been run.
  • The path from a trained model to a running engine (export, per-device compile, FP16 then INT8 after validation, a hardware-accelerated pipeline, per-camera profiling, versioned rollback) is designed. Earlier models ran on Jetson AGX Orin without that compile step.
  • Recognition on the airframe (Jetson Orin-class compute, an electro-optical and thermal pipeline, ROS 2 topics and MAVLink, gimbal cueing, plausibility checks, evidence on board, two-stage detect and confirm) is engineered to run and has not flown in that form.

Application areas without a model yet

Maritime and border recognition is described from market requirements gathered for maritime ISR programs. No model has been built for it.

What would change these statements

  1. A camera-domain evaluation, with per-class precision and recall and false alerts per camera-hour, on footage the models have not seen.
  2. A TensorRT build and benchmark of the deployed models on the target Jetson modules.
  3. Live camera ingestion and the on-site gateway, before any “real-time monitoring” claim.

When one of these lands, the site will say so, with the evaluation it rests on.

Built around your challenge

What do you need
to recognize?

Start a project