mean mAP
State-of-the-art mean average precision across the production model library, audited against the standard public benchmarks for every class the catalog covers.
| JAX · MOTION VIDEO · 150+ MODELS |
Jax for Detection. Pre-trained vision models across land, maritime, and aerial environments — composable with the agent, available as a tool, written back to the Codex as typed entities.
AGENT 02C · DETECTION
Motion-video computer vision used to be a destination. A separate console with its own model garden, its own console operators, its own queue of clips waiting to be hand-fed through the right pipeline. The detections came back hours later, in a format nothing else on the operator’s desk could read. Jax for Detection inverts that posture. The model library is part of the agent. The detections land in the same Codex everything else is already reading.
And this is what that changes:
Vehicles, personnel, structures, animals, equipment, signage. Trained on ground-level and oblique imagery across urban, rural, desert, and forest environments.
Vessels by class and size, wakes, oil slicks, debris fields, port activity. Tuned for daytime EO, low-light, and SAR-correlated tracks across littoral and open-ocean takes.
Fixed-wing, rotary, UAS, missiles, payload signatures. Operates on full-motion video and stills from airborne, tower, and orbital sensors.
Make, model, configuration, cargo state, load posture, license-plate region. Sub-models specialize in convoys, parking density, and traffic flow at scale.
Hull form, superstructure, mast configuration, weapon outline, container count. Fused with AIS when available, inferred when it isn’t.
Airframe type, livery, configuration, runway phase, mission posture. Includes a dedicated tail-number reader and a hangar-occupancy estimator.
Towers, pads, fuel depots, substations, pipelines, antenna arrays, construction stages. Built for monitoring change in fixed assets over long horizons.
A camera comes online. Jax pulls the stream, routes it to the right model in the catalog, runs the detection, and writes each result back to the Codex as a typed entity against the active ontology — coordinates, class, confidence, timestamp, source. From that moment on, every other surface treats the detections as first-class objects: the globe renders them, the map analyst can query them, the briefing tool can cite them.
WATCH LIVEmean mAP
State-of-the-art mean average precision across the production model library, audited against the standard public benchmarks for every class the catalog covers.
second latency
Sub-second per-frame inference on representative edge hardware. Cloud paths run faster on parallel takes; the agent chooses the path that fits the task budget.
inference modes
Edge, cloud, and hybrid — selectable per pipeline. The same model behaves identically across modes; only where it runs changes.
Every model in the catalog is versioned, signed, and benchmarked on the same fixtures every release. The active library is enumerated in app/(dashboard)/(home)/skills/_lib/model-catalog.ts — the same file the platform reads at runtime when the agent reaches for a model. New entries ship through the same review path as the rest of the codebase, with the benchmarks attached to the pull request.