A throwable sensor that classifies hundreds of acoustic signatures — vehicles, personnel, weapons, aircraft, machinery, wildlife, weather — and delivers plain-English threat assessments to the operator's tablet in under one second. In total darkness. Through smoke. Under full-spectrum jamming.
Built on the AudioSet taxonomy and extended with mission-specific signatures. The Puck doesn't listen for a handful of threats — it understands the entire acoustic environment.
Wheeled, tracked, engine type, speed estimation, approach vs. departure
Fixed-wing, multirotor, jet, propeller, including RF-silent platforms
Footsteps, speech, tools, digging, fence cutting, glass breaking
Gunshots, explosions, mortar launch, ricochet, caliber classification
Arcing faults, transformer hum, pipe stress, structural cracking
Wind, rain, wildlife, machinery — auto-profiled and suppressed from alerts
Each exists in isolation somewhere. No deployed system delivers all six. That's the gap we close.
The Puck emits nothing. No RF, no IR, no acoustic output. It cannot be detected, targeted, jammed, or spoofed. Total electromagnetic silence in contested threat terrain.
Two layers of AI run entirely on the sensor. Classification, reasoning, and alerting at the edge. When comms are denied, the Puck doesn't degrade — it was designed for that condition.
Not a narrow signature library. Full-spectrum acoustic classification across vehicles, aircraft, personnel, weapons, infrastructure, and environment. One sensor replaces an entire stack.
Not a code. Not a beep. An on-device LLM writes bearing, range, threat class, and confidence in natural language. Operators act on words, not waveforms.
Hockey-puck form factor. Battery-powered. 72 hours persistent overwatch. A team deploys a dozen across a square mile in minutes — no setup, no calibration, no babysitting.
Multiple pucks cross-reference bearings to triangulate threats in 3D. If a node goes down, the mesh adapts. If comms drop, detections cache locally. Resilient by architecture.
Permanent power, hardwired comms, remote server processing. No infrastructure, no capability. Useless forward of the wire.
Static signature libraries with a handful of preprogrammed sounds. Can't identify novel threats. Output a raw code an operator can't act on without an analyst in the loop.
Intercept radio signals only. Completely blind to fiber-optic munitions, dismounted personnel, vehicles running silent, gunfire, fence breaches, and every non-emitting threat class.
Push data to a remote server. Seconds to minutes of latency. Fail completely under jamming — the exact condition where threat detection is life or death.
Each Puck classifies hundreds of sound types independently. When multiple nodes detect the same event, the mesh triangulates and delivers a fused threat picture to the operator.
Two layers of on-device intelligence. Entirely passive. Entirely autonomous. Hundreds of classifiable sounds. Built for conditions where everything else fails.
A quantized neural network classifies 500+ sound types on-device in under 50 milliseconds. Vehicles, aircraft, personnel, weapons, machinery, wildlife, weather — the full acoustic environment, profiled and assessed before the echo fades.
When the classifier encounters ambiguity — overlapping sources, novel signatures, environmental masking — a compressed LLM reasons over the full acoustic context and writes a plain-English alert with bearing, range, assessment, and confidence. No analyst. No network.
Multiple pucks cross-reference direction-of-arrival bearings to triangulate sources in 3D. The mesh tracks vehicles, personnel, weapons fire, and aircraft simultaneously. Nodes that lose connectivity cache locally and sync on reconnect.
Detect approaching vehicles, dismounted patrols, weapons fire, and aerial threats — including RF-silent platforms — and push bearing + range to the operator's tablet in under one second. Works in total darkness, through smoke, under jamming.
Replace 24/7 manned observation with 72-hour autonomous acoustic overwatch. Track vehicles, personnel, fence breaches, and gunfire across the perimeter. Only confirmed events reach the operator — zero false-alarm fatigue.
Gunshot detection and localization for urban environments. Perimeter breach detection for critical facilities. Arcing fault detection for power infrastructure — hours before ignition. One mesh covers every acoustic threat domain.
No cameras. No RF. No cloud. No analysts. Just sound — and the intelligence to turn it into operator advantage.