Machine Learning Systems

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

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

Pure-Rust prompt-injection detector with 1.5MB embedded MLP classifier. 98.40% accuracy, p50 14ms CPU inference, bindings for Python/JS/Go. Apache-2.0/MIT alternative to Rebuff (archived) and Lakera Guard.

Ahead-of-time compiler that tiles ML models to fit 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)

Suite of edge-AI tools and a web dashboard that help schools cut water, food, and energy waste by detecting leaks, sorting compost, and visualizing consumption in one conclusive dashboard.

50.7x latency-optimized heterogeneous Vision Transformer (DeiT) hardware accelerator built on the Xilinx ZCU104 FPGA using DPU + custom HLS IPs.

Fajar Lang (fj) — Systems programming language for embedded ML & OS development. Compiler-enforced safety with @kernel/@device/@safe contexts. Rust-based compiler with Cranelift/LLVM backends. Made in Indonesia.

ESP32 camera that escalates from gentle reminders to airhorn if you slouch

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

Curated Edge AI resources for computer vision & audio: hardware, frameworks, benchmarks, literature, and communities (excluding mobile).

CS2 Skin Preview & Customization Utility for Weapons and Inventory is a visual tool for exploring and customizing weapon and inventory appearances in Counter-Strike 2, designed for previews, loadout styling, and cosmetic experimentation.

Real-time garbage sorting with MobileNetV3 on Raspberry Pi, STM32 handles the hardware side.

This is open source library for creating artificial neural network in c programming language for general purpose use.

Novel DBSCAN-based FPGA system for automatic modulation classification and NDA SNR estimation. O(n²)→O(n) complexity reduction. 71.7% lower power than state-of-the-art.

Notes and resources from Qualcomm On-device AI course, provided by DeepLearningAI

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

Estudo comparativo de arquiteturas de deep learning (CNN 1D, MLP, GRU, LSTM) para predição de temperatura em sistemas TinyML. Análise de performance, precisão e viabilidade para deploy em RP2040 com fusão de sensores AHT20/BMP280. Horizontes de 5, 10 e 15 minutos.

An energy-efficient TinyML fall detector that runs entirely on an Arduino Nano 33 BLE Sense. A cascaded pipeline keeps the CNN idle during everyday motion and runs it only on likely fall events, cutting per-event energy by about 98 percent versus an always-on model.