RAVEN 309 · Retail Theft AI
Raven 309 is our retail theft AI. The CAVE engine inside it — Concealment Action-Video Engine — is a movement, physics and behaviour-sequence intelligence layer that works with existing CCTV/NVR infrastructure, runs on a hidden edge host, and surfaces only honest, calibrated outputs to the operator.
Architecture
Every stage of the CAVE engine is deliberate. Detection, tracking, stabilisation, physics, neural reasoning and guards each have a defined contract, and only the final fused result reaches the operator.
Stage names above are deliberately abstract. Internal parameter sets, model choices and guard logic are part of the proprietary CAVE recipe inside Raven 309 and are not published. A calibrated decision threshold is applied at the Decision layer; only behaviour state and physics evidence are shown to the operator.
Pipeline detail
Raven 309 intentionally separates honest perception from learned reasoning. This is what keeps the operator output trustworthy.
The hidden edge host pulls live streams directly from the existing CCTV/NVR. Decode is hardware-accelerated, frames are aligned for synchronised overlay, and no NVR credentials are exposed beyond the edge.
The vision front-end produces a multi-point body skeleton in real time on the edge GPU. An identity tracker holds per-camera persistence across frames, and an adaptive stabiliser cleans the motion signal so the engine sees real movement rather than detector jitter.
The CAVE engine inside Raven 309 runs a deterministic physics block alongside a motion-memory buffer that captures kinematic signature over a short window. A sequence reasoner consumes both branches and feeds a set of calibrated decision heads. The internal model topology, parameters and guard logic are part of the proprietary CAVE recipe.
A sequential behaviour state machine combines with physics/context fusion and a stack of false-positive guards before the engine emits a single calibrated result. The exact guard composition and thresholds are tuned per release and are not published.
Engineering principles
Raven 309 was engineered against live retail CCTV constraints from day one: low-resolution streams, mixed lighting, partial occlusion, hoodies, and the long, sequential nature of real concealment. The CAVE engine inside it is not a re-skinned generic CCTV analytic.
Why we lead
Generic CCTV analytics rely on box-level motion detection or single-frame classifiers. Raven 309, with the CAVE engine inside, was built specifically for concealment, against the kinds of cameras retailers actually run, with a refusal to display anything the model was not trained to say.
A behaviour state machine requires a real temporal pattern before any alert is fused. One-frame coincidences cannot pass.
Deterministic posture, reach and locomotion cues run alongside the neural reasoner. Even if one branch drifts, the other keeps the operator honest.
CAVE is movement-driven. It does not profile faces, demographics or attire. It looks at how a body moves, not who it belongs to.
Only the engine’s calibrated, trained outputs reach the operator. No fabricated scores, no unsupervised “intent” verdicts — ever.
The AI runs on a sealed edge host. Live overlay stays inside the shop LAN; only alerts, clips and health flow to the customer portal.
A universal baseline deploys first; controlled offline calibration produces store-specific tuning when a site needs it. Recipe stays in-house.