SCADAgric
LLM function-calling framework for agricultural IoT sensors at the edge
SCADAgric is an edge-native framework that enables large language models to directly call functions on real-time sensor streams — without routing data through the cloud. Built for precision agriculture, it demonstrates how LLMs can serve as autonomous reasoning agents in low-connectivity environments.
The Problem
Agricultural sensors (soil moisture, temperature, humidity, CO₂) produce continuous time-series data. Existing IoT deployments require cloud round-trips for any intelligent decision-making — untenable where network connectivity is intermittent or expensive, as in much of Sub-Saharan Africa.
What We Built
- A function-calling interface that exposes sensor read, threshold-check, and actuator-trigger operations as typed tool definitions consumable by an LLM
- An edge inference runtime running on Raspberry Pi / NVIDIA Jetson that processes LLM-generated function calls locally
- A semantic query layer: farmers or agronomists can ask natural-language questions (“Is soil moisture in field 3 below the irrigation threshold?”) and receive grounded answers from live sensor data
Key Outcomes
- Zero cloud dependency for real-time sensor queries — all inference runs locally
- Demonstrated on a 12-sensor prototype greenhouse setup
- Latency under 200 ms for sensor read + LLM reasoning cycle on Jetson Nano