embedded-ml

Machine Learning Systems: Foundations, Scaling, Agentic AI, and Physical AI (Vols I–IV) • Harvard CS249r | https://mlsysbook.ai

TinyML & Edge AI: On-device inference, model quantization, embedded ML, ultra-low-power AI for microcontrollers and IoT devices.

Ahead-of-time compiler that fits ML models on embedded devices with hard memory budgets.

Python ML for training a custom on-device cry model (knowledge-distilled from YAMNet, INT8, deployed on ESP32-S3)

Hardware-aware face detection on Samsung GT-S7392 (ARM Cortex-A9)

Static INT8 AOT compiler that turns fixed PyTorch MCU models into standalone C11

ChatTLM: a 10.9M-parameter language model that runs on a TI-Nspire CX II CAS graphing calculator, with the code, data and measurements behind the paper "Letting the Tools Do the Math"

Six-class activity classifier with compact softmax modeling, int8 quantized C++ inference, reproducible benchmarking, and an ESP32 replay sketch using UCI HAR data.

Compress PyTorch models for edge devices — CPU-only, no GPU, no retraining. One function call.

Early-exit, full-model and static-baseline inference on RubikPi 3, with live-camera, HTP, latency and thermal evidence.

Don't Think It Twice: Exploit Shift Invariance for Efficient Online Streaming Inference of CNNs

Wearable fall-detection device with double-layer LSTM (97.8% accuracy), companion to published research

Deploy and manage ML models at the edge — OPC-UA integration, PLC connectivity, real-time inference on embedded hardware for sub-millisecond decisions

Cloud-to-edge image classification: ResNet-18 fine-tuned in Colab, exported to ONNX, deployed to an NVIDIA Jetson Nano for on-device inference. NVIDIA DLI Jetson AI Specialist final project.

A lightweight benchmark for tiny ML models on ESP32-class devices.

RIN Engine Runtime - Universal Inference Engine

Siren direction-of-arrival on Infineon PSOC Edge: 4-mic DS-CNN, int8, Ethos-U55 (EESTech hackathon)

Offline Sri Lankan Sign Language recognition on Raspberry Pi 5 — MediaPipe hand landmarks + RandomForest, streamed to a Flask dashboard. No cloud, no internet.

Multiposition heart sound analysis

Satellite-image CNN compressed 7.5x to a 0.83 MB INT8 ONNX model: 95.7% top-1, 0.76 ms p95 on one CPU core, and 95.9% retained under simulated radiation bit-flips that reduce FP32 to random guessing.