14–15 Nov 2026
Bengaluru
Asia/Kolkata timezone

LLM on a $10 microcontroller: natural-language hardware control with no cloud

15 Nov 2026, 14:10
30m
Audi 1

Audi 1

Talk (30 mins) Embedded Systems & IoT

Speaker

Mr Manjunath Janardhan

Description

Most "on-device AI" is a thin client talking to someone else's GPU. This talk is the other end of the spectrum: a transformer trained from scratch, small enough to run on a $10 microcontroller, that turns English into GPIO commands with no network of any kind. No WiFi, no cloud, no API key — there is no network stack in the firmware at all.

You type "blink pin 4 every 5 seconds 3 times" at a serial prompt and an LED blinks. The model is 312,128 parameters of fp32 weights memory-mapped out of flash, running on an ESP32-S3 next to the code that drives the pins. There is no operating system on the device: the model, the tokenizer and the GPIO driver are one firmware image. Training takes about five minutes per run on a laptop, and the whole toolchain is Linux-native.

The interesting part is not that it fits. It is what went wrong.

I gave the model a grammar in which an invalid request was impossible to express
— every legal pin had its own token, so it physically could not say "pin 100."
That reads like a safety guarantee. It is the reverse. A model that cannot say "pin 100" does not refuse; it says the nearest thing it can. On held-out data,
"switch off pin 100" came back as pin 10: a real pin, a valid command, actuated silently, with nothing in the logs.

That failure mode is not specific to microcontrollers. Constrained decoding, GBNF grammars, JSON-schema enforcement and function-calling APIs all work by restricting what a model may emit and then sampling the most likely permitted token. If your schema cannot express "the input was out of range," the model returns something in range and you never see the difference.

The fix inverts the responsibility: numbers are transcribed digit by digit so the model can say something invalid, and the hardware layer refuses it by name.

Speaker Bio

Manjunath Janardhan is an AI/ML Computational Science Senior Manager at Accenture with over 22 years of experience in enterprise software, healthcare technology, cloud platforms, and Artificial Intelligence. He specializes in Agentic

AI, multi-agent systems, LLM applications, AI governance, and cloud-native AI platforms. Prior to Accenture, he served as Principal AI Engineer at msg global solutions and spent nearly 13 years at GE Healthcare, where he architected GenAI platforms, intelligent service solutions, and enterprise software systems. An active open-source contributor, technical reviewer, mentor, and technology writer,

Manjunath regularly shares insights on emerging AI technologies, LLM
infrastructure, and enterprise AI adoption.

Prerequisites

None required. The hardware is explained as it appears, and the central idea needs no machine learning background.

If you have used structured outputs, JSON mode, function calling or
grammar-constrained decoding, the main finding will apply directly to work you are already doing.

Summary

A transformer trained from scratch, small enough to run on a $10 microcontroller, turns English into GPIO commands with no WiFi, no cloud and no API key. Along the way it showed that making invalid outputs impossible to express does not make a model refuse them — it makes it substitute a valid one, silently. Live demo on real hardware, with the failures included.

Difficulty Intermediate
Speaker LinkedIn / Socials https://www.linkedin.com/in/manjunath-janardhan-54a5537/

Presentation materials