tiny-ml

Resources of our survey paper "Optimizing Edge AI: A Comprehensive Survey on Data, Model, and System Strategies"

MINERVA - Minimal Inference Engine for Robust, Verifiable, and Authenticated ML. Encrypted, integrity-verified neural network inference for MCUs down to ATmega328P.

[ICLR2022] Efficient Split-Mix federated learning for in-situ model customization during both training and testing time

Benchmarking machine learning inferencing on embedded hardware.

Source code of “Agile Reinforcement Learning for Real-Time Task Scheduling in Edge Computing” (CAIAC 2025)

An editable, auditable 807K-param byte-level LLM: CRUD single facts with provable per-edit locality, and abstain when unsure instead of guessing. CPU, offline.

Epsilon is a library with functions for machine learning and statistics written in plain C. It is intended to run on microcontrollers.

Ultra-low-latency Ternary Neural Network inference engine for HFT alpha generation, optimized for an ultra-budget 1,120-LUT Renesas FPGA. Features a fully unrolled combinatorial popcount tree achieving 2-cycle deterministic execution without hardware multipliers, paired with an ESP32-S3 hardware-software co-design.

Central Proteus simulation and Git submodule tracker for the STM32 Smart Car network.

This project focuses on the implementation of optimized Linear and DNN regression models for inter-vehicle distance prediction in a Cooperative Adaptive Cruise Control (CACC) application. It leverages Tensorflow Lite to create optimized models through quantization and pruning for realtime inferencing on Raspberry Pi and On-board Unit (OBU) of Co…

TinyML-based obstacle detection and emergency braking system simulated on STM32F401VE using STM32Cube.AI, CMSIS-NN, and Proteus. ONNX model deployed on Arm Cortex-M4.

Implementación de un clasificador de vocales basado en SVMs y features MFCC.

PXG-Tiny Ai - pixel sprites game asset generator. A 483K-param open-weights model that turns plain English prompts into 16×16 pixel-art sprites — ~1 MB NumPy weights, CPU-only, fully offline.

Real-time wildfire detection on microcontrollers with TinyML + Edge Impulse + Wokwi

A fast, smooth, algebraic alternative to Mish for TinyML and Edge AI. LiteMish uses only addition and multiplication, resulting in improved computational efficiency while preserving learning performance.

A Smart Mask Enforcement System using Multitenant Cascading Architecture in TinyML

Tiny CNN cat/dog classifier for RISC-V edge AI, with PyTorch training, quantized firmware export, and Renode simulation benchmarks.

A new communication paradigm proposal for restricted bandwidth and fragile channel conditions in NTN.

Tiny open-source model (1.2 MB) that understands percentage questions in English, Portuguese and Spanish and returns the right calculation. Runs in the browser in ~1 ms.

Edge-optimized wildlife detection for crop protection