Architecture

How Sentinel Works

Capture → infer → encrypt → tethered review. Four stages, one device, zero wireless biometric egress.

Capture
5 cams · IR · thermal · mics
Inference
On-device models
Encrypt
AES-256 enclave
Tethered review
USB-C + app

1 · Multimodal Capture

Five UHD cameras (6–8 MP) capture macro and micro body language. IR captures pupil dilation and low-light gaze. A non-contact thermal sensor reads skin temperature. A dual beamforming mic array isolates the subject's voice from the room.

5× UHDIR / gazeThermalArray mics

2 · On-Device Vision & Audio AI

Snapdragon 8 Gen 4 + dedicated GPU runs the inference stack: facial landmark + AU detection, micro-expression classification (FACS-based), posture & gesture estimation, vocal stress and prosody analysis. Streams are time-aligned and fused into a single behavioral signal track.

SD8 Gen 4FACS · pose · prosodySensor fusion

3 · Hardware-Encrypted Enclave

Raw frames are never written unencrypted. AES-256-GCM encryption is performed inline by the SoC's secure enclave; keys are wrapped by a device-bound master key that is generated at first boot and cannot be exported. Pulling the storage chip yields ciphertext only.

AES-256-GCMHardware-bound keys

4 · USB-C Tethered Review

Sentinel has no wireless data egress for biometric records. The only way to read what it saw is the Sentinel companion app, on your laptop, with the device physically plugged in via USB-C. Mutual attestation verifies device + app + operator before any data is unlocked.

USB-C onlyMutual attestation

Privacy posture

Sentinel ships with capture indicators that cannot be disabled in firmware. Subjects are notified by a visible recording light on the front bezel. Sessions auto-expire; raw footage is purged from the enclave after operator-defined retention windows. Factory-reset pinhole destroys the master key, rendering all stored data permanently unreadable.