A full-stack IoT architecture for real-time detection, analysis and reporting of seismic events. Everyday household appliances — washing machines, TVs, refrigerators — become a distributed earthquake early-warning network that alerts you before the shaking arrives. v1.2.0 AI generates emergency reports on-premise with Ollama — privacy-first, seismic telemetry never leaves the machine.
We want to make earthquake detection available all around the world, with zero cost for the people and low cost for the enterprises. By embedding miniature IoT sensors inside the objects we already live with, we can grant P-wave detection without anyone even knowing — and give people time to evacuate or reach a safe location before it is too late.
A faithful, client-only echo of the backend's POST /demo/trigger-earthquake flow:
pick one of the 8 geographic zones, set a magnitude ≥ 4.5 and watch the pipeline run from edge to mobile alert + AI report.
Historically the backend bypasses the IoT pipeline and pushes to the
Redis quake_alerts channel — the mobile app vibrates and fires a push notification within milliseconds.
Moderate-strong event detected by 12 sensors in Sicily. Shaking intensity IV-V MMI; recommend staying away from windows and tall furniture, and monitoring aftershocks over the next 2 hours.
quakeguard/telemetry on HiveMQ Cloud (TLS, port 8883).POST /readings/. API key, sensor status, anti-replay (60 s) and signature are all validated.202 Accepted instantly.M = log10(PGA × 1.6) + 3.0; if M ≥ 4.5 an Alert is triggered.ai_report_queue in Redis → worker → Ollama generates the emergency report on-premise; persisted and pushed on the ai_reports WebSocket channel.ESP32-C3 (RISC-V) + ADXL345 accelerometer. Runs a FreeRTOS DSP pipeline with O(1) ring-buffer STA/LTA and ECDSA signing on every payload.
ESP32-C3STA/LTA100 Hz
Fully async Python 3.11 gateway. Sliding-window rate limiting, API-key auth, ECDSA verification and anti-replay — designed to sustain the Thundering Herd.
AsyncRate limitUvicorn
Decouples ingestion from processing (producer-consumer). Queue offloading, 60 s per-zone dedup locks and a Dead Letter Queue on failure — plus a dedicated ai_report_queue for the ai_report_worker.py, decoupled from the alert engine.
HiveMQ Cloud broker over TLS. A dedicated bridge microservice forwards telemetry to the secure HTTP pipeline, injecting the API key.
TLS 8883PahoBridge
Spatial persistence. Sensors auto-assigned to the smallest containing macro-region via ST_Contains, ordered by area.
Three-tab app (Monitor, Sensor Map, Settings). WebSocket alerts with SOS haptics & push, live seismograph, offline mode, last-10 alert history, plus a new AI Report UI: banner for the latest alert, feed cards with summary + recommendations, and a "AI Report Unavailable" badge on FAILED.
ExpoZustandTanStack Query
Earthquakes don’t knock on one door. When hundreds of sensors in the same zone detect the same event at once, the backend is hit by a massive, synchronous traffic spike.
QuakeGuard is engineered for exactly that moment: ingestion is decoupled from processing, requests return
202 Accepted in milliseconds, connection pooling absorbs the load, and a
per-zone Redis cooldown guarantees the network fires one clean alert — not a storm of duplicates.
Ingestion never waits on the database; events offload to Redis instantly.
50 req/s per IP, backs off buggy or malicious nodes early.
Aggressive SQLAlchemy pool configured for high concurrency, no queue exhaustion.
Dedup lock per zone; only the winning worker publishes the Alert (outbox pattern).
Data integrity is paramount in an emergency system. Every telemetry packet is cryptographically secured end-to-end against spoofing, replay and tampering.
Checked on every request using a constant-time comparison to resist timing attacks.
Confirms the sensor exists and is marked active in the database.
Timestamps older than the 60 s window are rejected outright with 403 Forbidden.
Payload verified against the device public key — both DER and raw r||s forms supported.
On first boot each ESP32-C3 generates its own NIST P-256 ECDSA keypair (mbedTLS).
The private key is sealed in Non-Volatile Storage and never leaves the device; the public key becomes the
unforgeable identity used during the automated /devices/register provisioning handshake.
8 global macro-regions are pre-seeded. Sensors are auto-assigned to the smallest containing polygon via a PostGIS spatial query — with an Unknown Region fallback for unmapped coordinates.
Edge seismic detection on ESP32-C3 and local alerts.
HiveMQ Cloud MQTT (TLS), ngrok HTTPS tunnel, security hardening.
LLM backend for automatic emergency-report generation from MQTT telemetry.
Accurate node timestamps, GPS coordinate resolution and ADXL345 calibration.
Multi-node spatial correlation + AI reports for epicenter calculation.
Grafana live visualization of seismic telemetry.
Five automated pipelines gate every change:
backend-ci — Bandit, Safety, stress testfrontend-ci — ESLint, npm auditiot-ci — PlatformIO compilationpr-lint + devops-ciValidates the full ingestion → Redis → worker → PostGIS → WebSocket pipeline against a simulated massive seismic event.
A Typst-compiled whitepaper walks through architecture, hardware, security, the broker, backend, mobile and deployment.
Open source under AGPL-3.0, with a DOI, contributing guide and security policy.
Developed as a school-contest project for Hackersgen by Sorint.lab and the GF Marilli competition — in collaboration with riccardo0731.
We were honored to work with Francesco Finazzi — UniBG professor and founder of Earthquake Network, the largest community seismic network in the world. After the conference, our collaboration with him is exactly where the project’s story began.
Sorint.lab — Hackersgen
University of Bergamo
The demo video is currently available only in Italian.
QuakeGuard is open source and community-driven. Ideas, issues and pull requests are always welcome — pull in, and let’s build a safer world together.