After a rockfall, engineers can reconstruct exactly what happened — the displacement, the trigger, the failure sequence. That analysis is precise, thorough, and almost worthless to the people who were standing at the bottom of the slope.
The entire value in hazard detection lives on the front of the event, not the back. That's the problem I built an AI + IoT rockfall prediction system to attack — and it was shortlisted at Smart India Hackathon 2025.
Why one sensor can't save you
The instinct is to slap a displacement sensor on a slope and set a threshold. It doesn't work, because no single sensor is trustworthy on its own. Displacement, vibration, moisture — each is noisy, each throws false alarms, and a system that cries wolf gets ignored right before the one time it's real.
The signal isn't in any one stream. It's in how they move together.
Fusing the streams
The system runs a multi-sensor data-fusion pipeline feeding a real-time hazard model that weighs the sensors jointly. When you combine them, the noise starts to cancel and a genuine pattern rises out: this combination, trending this way, means the slope is about to move — evacuate now, not after. On the held-out evaluation it reached a 98% AUC-ROC.
Two engineering priorities shaped it:
- Fusion over thresholds. Instead of independent trip-wires, the model reasons over the sensors as a system, which is what separates a real warning from a jittery false alarm.
- The edge, because latency is safety. Inference runs close to where the sensors are. A hazard warning that depends on a cloud round-trip is a hazard warning you can't rely on when connectivity fails — which, on a remote slope, is exactly when you need it.
The mindset
Prevention systems are judged by their lead time, not their hindsight. A crude alarm a minute early beats a flawless post-mortem an hour late, every single time. Designing for the warning — for the moment before — changes every decision: which sensors, where the compute lives, how you tune the trade-off between false alarms and missed events.
Building this reframed how I think about ML entirely: the accuracy number matters, but when the answer arrives matters more. The full pipeline and architecture are on the project page.
👉 See the system: www.divyakush.com/projects/rockfall-prediction
Divyakush Punjabi — Full-Stack & AI Systems Engineer
🌐 https://www.divyakush.com · 💼 LinkedIn · 💻 GitHub
Top comments (0)