Bridging browser-based WebAI and microcontroller TinyML through transparent, single-file implementations.

Welcome to WebMCU-AI Lab. Led by a veteran Technology & Robotics educator with over 35 years of experience, this organization focuses on demystifying "Black Box" AI through Transparent TinyML and WebAI integration — from a $15 microcontroller to a Chrome browser, no cloud required.

We specialize in On-Device Machine Learning (TinyML) with browser-based interaction via WebSerial and TensorFlow.js. Our work spans the full training spectrum:

• Fully on-device — complete CNN training and inference on ESP32-S3 with no external computation
• Browser-assisted — TensorFlow.js trains in Chrome, weights transfer to the MCU via WebSerial
• Hybrid — on-device fine-tuning of browser-trained weights

All implementations are single-file, dependency-free, and designed to make every step of the ML pipeline visible and modifiable.

• 🔍 Transparent AI — no black boxes; every weight, gradient, and activation is accessible and documented
• 🎓 Inquiry-Path Pedagogy — foundational numeracy and logic before tool-based abstraction; designed for K-12 through undergraduate research
• ⚡ Edge-to-Web Integration — seamless WebSerial communication between microcontrollers (ESP32-S3, Arduino Portenta H7) and browser interfaces
• 🔋 Energy Transparency — direct measurement of complete ML pipeline energy footprint, including training and inference

• "On-Device Vision Training, Deployment, and Inference on a Thumb-Sized Microcontroller"
  Submitted to WCCI 2026 (under review)
  Complete CNN backpropagation on ESP32-S3 — 1,750 lines of C++, no cloud, no external dependencies.

All lab repositories follow these standards for clarity and reproducibility:

• Vanilla Everything — minimize dependencies; prioritize single-file HTML and vanilla JS with inline CSS
• Naming Convention — descriptive camelCase with a my prefix for all internal functions
  (e.g., async function myTrainModel())
• Logic First — async/await over .then() promises for readable, linear control flow
• Single-File Firmware — complete Arduino sketches in one .ino file; every component visible without navigating a library tree
• MIT License — all code open source and freely reusable

• Seeedstudio XIAO ESP32-S3 Sense — primary platform
  ($15-40 USD, 8MB PSRAM, OV2640 camera, touch, OLED)
• Arduino Portenta H7 — secondary platform
  (ports in progress)

We collaborate with industry and academic researchers to bring Edge AI into the K-12 and hobbyist ecosystem.

Member: TinyML4D / MLSys Community

Contributions, ports to new hardware, and curriculum adaptations are welcome. See individual repository CONTRIBUTING.md files.

• GitHub: @hpssjellis
• LinkedIn: Jeremy Ellis

"Master the logic, then command the machine."