Inspiration
When natural disasters strike, cellular towers and power grids collapse in seconds. Trapped survivors tap on concrete beams or call out in regional dialects. But in a total blackout, emergency response cannot rely on cloud AI, nor can teams afford false alarms triggered by falling debris, rain, or heavy recovery machinery. A conventional threshold detector can easily be forced into a binary guess on ambiguous signals. We asked: Can we build an offline intelligence network that doesn't just detect events, but mathematically determines whether the physical evidence is sufficient to act?
What It Does
SignalGate is a local-first emergency triage network combining framework-light acoustic sensing, multi-dialect crisis keyword spotting, and mathematical evidence arbitration completely offline:
- Framework-Light Acoustic Ingestion: A zero-dependency pure-NumPy circular buffer detects rhythmic distress tapping (Morse SOS) in real time (measured at 0.14 ms on our Windows 11 test machine) without heavy audio frameworks.
- Multi-Dialect Crisis Keyword Spotting: An offline crisis distress lexicon fixture structured across 5 vulnerable regional disaster dialects (Tagalog, Haitian Creole, Ukrainian, Turkish, Spanish).
- Evidential Triage Arbitration: Applying Dempster-Shafer Theory of Evidence over Frame $\Omega = {\text{Victim}, \text{Noise}, \text{Hazard}}$, SignalGate measures Epistemic Ignorance ($m(\Theta)$) and Inter-Sensor Conflict ($K$). When signals conflict, the system actively triggers
HOLD_AND_CORROBORATEinstead of an expensive false alarm. - Decentralized Corroboration: When neighboring field nodes provide independent spatial corroboration, the system escalates to an authenticated
P0 DISPATCH, routed across simulated decentralized mesh hops with cryptographic HMAC-SHA256 provenance receipts.
How We Built It
- DSP Engine: Built from scratch in pure Python/NumPy with circular ring buffers and sliding FFTs.
- Evidence Arbitration: Implemented Dempster's rule of combination to compute joint belief and epistemic ignorance in 0.01 ms.
- Chaos Mode: A built-in fault injector subjecting the network to 40% packet loss, -5dB acoustic noise, and contradictory sensors.
- Test Integrity: 15 automated unit tests verifying DSP edge cases, ignorance bounds, conflict refusal, and HMAC tamper detection.
Challenges & Breakthroughs
Moving from naive classification to uncertainty-aware arbitration. In our 50-scenario synthetic Chaos Mode fault suite, SignalGate achieved 0 false dispatches out of 25 high-conflict scenarios on our test machine by refusing to escalate uncorroborated noise.
What's Next
Porting the pure-NumPy DSP algorithms to low-cost microcontrollers for humanitarian field deployment.