From raw RF to actionable intelligence — entirely on the sensor, entirely in real time. A purpose-built stack for edge AI inference in denied, degraded, and disconnected environments.
DC–6 GHz
6.144 GSps
16-bit resolution
Channelization
DDC / DUC
FFT engine
INT8 quantized
<2 ms latency
10k+ signal types
Multi-node mesh
Sub-meter accuracy
Real-time tracking
REST / gRPC API
TAK integration
ATAK / WinTAK
Total pipeline latency from antenna to classified alert: <2ms on SWaP-optimized hardware. Zero data leaves the node unless explicitly commanded.
Strangeloop's inference engine is built on a custom neural architecture trained on the world's most comprehensive RF signal dataset — over 4 terabytes of annotated captures spanning commercial, government, and adversarial emitters. Models are quantized to INT8 and run entirely on-device. No GPU cluster, no cloud endpoint, no network dependency.
Temporal convolutional networks (TCN) with attention mechanisms. Trained on IQ samples, not demodulated bits.
On-device few-shot adaptation to novel waveforms. New signal profiles deploy via signed OTA update.
Every classification carries a calibrated confidence score. Low-confidence detections trigger expanded collection.
Models trained with adversarial examples to resist intentional waveform manipulation and evasion.
Time-difference-of-arrival across a mesh of Strangeloop nodes. Nanosecond-precision timestamping with hardware-level PPS sync.
Angle-of-arrival via phased-array antenna processing. Single-node bearing fix in under 500 μs. Multi-node triangulation for 3D position.
Kalman-filter fusion of TDOA, AoA, and RSSI measurements. Adapts dynamically to node availability and SNR conditions. Sub-meter CEP50 at 2 km.
| Range | CEP50 | Nodes Required |
|---|---|---|
| 500 m | <0.3 m | 3 |
| 2 km | <0.8 m | 3 |
| 10 km | <5 m | 4 |
| 50 km | <50 m | 4+ |
| PPS source | GPS / GNSS or disciplined oscillator |
| TDOA resolution | 1 ns |
| Sync error budget | <5 ns RMS |
| GPS-denied fallback | Disciplined TCXO, <50 ns/hr drift |
Strangeloop's monitoring engine doesn't require an operator to specify what to look for. It continuously maintains a spectral baseline, detects deviations, and autonomously adjusts scan parameters to track threats.
Priority queuing ensures the most operationally relevant signals receive maximum dwell time — without dropping awareness of the broader electromagnetic environment.
Extended frequency ranges available via optional front-end modules.
| Instantaneous BW | 6.144 MHz – 6 GHz (multi-tuner) |
| Scan rate (full band) | <10 ms per sweep |
| Minimum dwell | 1 µs per channel |
| Concurrent monitors | Unlimited (software-defined) |