Everyone builds ML like the model lives in a datacenter — abundant compute, reliable network, all the RAM you want. Then you try to put intelligence on a remote mountain slope with no reliable connectivity, and every one of those assumptions collapses. That's where edge ML gets real.
I ran straight into this building an AI + IoT rockfall prediction system, and it changed how I think about where inference should live.
The assumption that breaks
Cloud ML is seductive because it hides the hard parts. Send data up, get a prediction down, don't worry about the compute. But that model quietly assumes the network is always there — and for anything monitoring the physical world in a remote or hazardous place, the network is exactly what you can't count on. The moment you most need a prediction is often the moment connectivity is worst.
If your safety system depends on a round-trip to a server, it isn't a safety system. It's a demo that works when everything's fine.
Pushing inference to the edge
So the intelligence runs close to the sensors. A multi-sensor fusion pipeline processes displacement, vibration, and moisture streams locally and feeds a real-time hazard model that reached 98% AUC-ROC — the decision happens where the data is born, not a datacenter away. The system was shortlisted at Smart India Hackathon 2025.
The edge forces a different discipline:
- Latency is safety, not a metric. A warning that arrives after a cloud round-trip is a warning you can't stake lives on. Local inference closes the loop immediately.
- Resilience by default. When the compute lives at the edge, a dropped connection degrades reporting, not detection. The slope keeps being watched whether or not the uplink is up.
- Do more with less. Constrained edge hardware means every design choice about the model and pipeline has to earn its keep — which, honestly, makes for better engineering than infinite cloud resources ever do.
The takeaway
"Where should the inference run?" is one of the most underrated questions in applied ML. For anything with a physical, time-critical consequence, the answer is usually: as close to the sensor as you can get. The cloud is still useful — for the dashboards, the history, the retraining — just not for the reflex.
Building for the edge taught me that the constraint isn't a limitation. It's the thing that makes the system trustworthy. The full pipeline is 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
