Every time I've seen a project need real AI reasoning at the edge, in industry, in the field, the answer is almost always the same: bring in an NVIDIA Jetson, or something in that same class. Powerful, capable, and expensive. Every single time.
It's treated as a fixed rule. If a device needs intelligence, it needs expensive hardware.I set out to check whether that rule actually held, or whether it just hadn't been tested.
For my final year research, I set out to answer a simple question: could a $10 microcontroller, with no AI model running on it and no hardcoded logic, still make genuinely adaptive, intelligent decisions? Not simulated intelligence. Not a lookup table dressed up as smart. Real reasoning, applied to a real problem.
I chose irrigation as the test case, since it's a problem that exposes the limits of simple automation. A fixed timer waters a cactus the same way it waters a tomato. A threshold sensor doesn't distinguish between a plant in its first week and one about to flower. Effective irrigation requires judgment: species, soil, weather, and growth stage all considered together.
So I built a system where the microcontroller, an ESP32-S3, does almost nothing intelligent at all. It reads soil moisture, temperature, and humidity. That's it. No model on the chip. No plant logic in the firmware. It has no awareness that AI is involved anywhere in the loop.
All the actual reasoning happens elsewhere, in the cloud, across seven specialized LLM agents coordinated through a shared database. One agent identifies the plant's requirements. Another reads the weather. Another calculates the exact water requirement using FAO-56 agronomic models. Before any decision is allowed near real hardware, it must pass a 10-point safety gate designed to catch invalid or unsafe commands before they reach the device.
Then came the part I was most curious about: does the choice of LLM actually matter here?
I ran the entire system end to end, 65 times, comparing two models as the reasoning core. Nemotron-3 Super 120B completed 100% of its runs with 96.4% decision accuracy. GPT-OSS-120B, despite having the same parameter count, failed every single run.
The reason why was the most interesting finding of the project. It wasn't a knowledge gap. It was an inability to reliably chain tool calls across multiple agents in sequence, a distinct skill from raw model size, and one that mattered far more than I expected.
The result is a fully working, hardware-validated system: a $10 chip participating in real, adaptive, AI-driven decision-making, with no Jetson, no dedicated AI hardware, and no AI running on the device itself.
This research was accepted at WUETecS 2026, where I presented the poster.
Grateful to my supervisor, Mr. T.M.P. Tennakoon, for his guidance throughout this project.
Full repo - https://lnkd.in/gqsjdG_w
#EmbeddedSystems #ArtificialIntelligence #LLM #IoT #EdgeAI #MultiAgentSystems #ESP32 #AgriTech #Research #EmbeddedAI #SmartIrrigation
1 August 2026 à 09h23
View on LinkedIn →
