EchoShield

Inspiration

During active shooter events, human-in-the-loop 911 reporting creates fatal communication latency. First responders need immediate, verifiable intelligence, not chaotic phone calls. EchoShield eliminates human reaction time by transforming everyday mobile phones into a decentralized, zero-latency acoustic sensor mesh. It is the only system that provides autonomous threat detection and mathematical ground-truth corroboration entirely on edge devices, giving enterprise and academic campuses an invisible, instantaneous defense infrastructure.

What it does

EchoShield is a scalable, privacy-first mobile sensor mesh designed for active threat environments. It bypasses human intervention by executing machine learning directly on edge devices to identify acoustic anomalies, coordinate multi-node corroboration, and dispatch structured intel to authorities.

Privacy-Isolated Edge Inference: Constantly analyzes ambient audio via hardware-accelerated, quantized classifiers. Raw voice data never leaves the device's volatile memory, ensuring zero cloud interception.
P2P Threat Consensus: Devices dynamically form a localized mesh to cross-validate acoustic signatures mathematically, neutralizing false positives before initiating a network broadcast.
Agentic State Summarization: An autonomous LLM agent ingests corroborated multi-node telemetry to compile structured, police-ready dispatch briefs in real-time.

How we built it

We architected EchoShield across three strict technical domains to eliminate single points of failure and guarantee deterministic execution.

The Edge

Asynchronous Audio Gating: We engineered a two-tier acoustic pipeline to preserve battery. A low-power amplitude "tripwire" monitors ambient noise, only waking the heavy inference threads when decibel thresholds breach a pre-calculated anomaly baseline.
Zetic MLange on NPU: We deployed an 8-bit quantized convolutional neural network directly to the Neural Processing Unit using Zetic MLange. This executes the heavy classification instantly without starving the primary CPU or compromising thermal limits.

The Mesh

Wi-Fi Aware Local Topology: We bypassed fragile cellular networks entirely. Leveraging Google Nearby Connections and Wi-Fi Aware, devices construct an ad-hoc P2P_CLUSTER that survives internet blackouts and tower congestion.
Sentinel Duty Rotation: To mitigate fleet-wide battery drain, the cluster dynamically elects a rotating listener node. Upon anomaly detection, the sentinel broadcasts a localized WAKE:CLASSIFY packet, forcing surrounding nodes to execute their local models and formulate a P2P Consensus.

The Cloud

High-Throughput Node.js Relay: A zero-dependency intake layer aggregates localized mesh decisions into a global state array. It actively strips duplicate messageIds and drops stale packets to maintain a clean operational picture.
LLM Dispatch Coordinator: A custom AI agent parses the incoming mesh state tensor, maps node coordinate drift, and generates real-time operational intelligence for first responders.

Challenges we ran into

Continuous Buffer Hardware Sandboxing: Modern mobile OS architectures aggressively sandbox background audio. We circumvented OS-level hibernation by engineering our custom two-tier audio gate and the "Sentinel" peer rotation algorithm, slashing power draw while maintaining a strict sub-200ms detection threshold.
Mesh Broadcast Storms & Packet Collisions: During a simulated threat threshold, concurrent node broadcasts saturated the local network band. We mitigated these collisions by implementing strict messageId deduplication caches and packet Time-To-Live (TTL) constraints, preventing infinite broadcast loops.
LLM Telemetry Hallucinations: Translating chaotic, asynchronous node states into a structured dispatch brief induced severe LLM hallucination. We stabilized the pipeline by injecting a strict JSON-schema enforcement layer and deploying mathematical penalty prompts to suppress unverified state assumptions.

Accomplishments that we're proud of

Sub-Second Critical Machine Loop: We successfully engineered a highly deterministic end-to-end pipeline: acoustic tripwire → Zetic MLange inference → P2P Consensus → AI dispatch generation → encrypted UI push alert, all executing within milliseconds.
Zero-Trust Edge Compute: We proved the viability of deploying persistent, heavy acoustic models on consumer-grade hardware. We deliver military-grade campus defense without ever transmitting raw ambient audio over an external network.

What we learned

NPU Memory Alignment: Compiling architectures for mobile NPUs requires exacting tensor shape management and memory alignment. The engineering overhead is steep, but the resulting thermal and energy efficiency vastly outperforms standard CPU fallback.
Decentralized State Resolution: We learned that establishing ground truth in a headless P2P_CLUSTER requires rigorous algorithmic consensus. Simple packet broadcasts are insufficient; true mesh reliability demands decentralized mathematical agreement to filter out localized noise anomalies.

What's next for EchoShield

Enterprise Headless SDK: Abstracting our core mesh and inference architecture into a white-labeled SDK. This enables zero-friction deployment into existing enterprise security applications and campus safety suites.
Legacy CAD Integration: Developing dedicated API bridges to inject our structured operational telemetry directly into legacy 911 Computer-Aided Dispatch (CAD) systems.