Project Diya is a wearable AI device that sees obstacles, faces, and currency for its user, while guiding them with voice and haptic cues.
An experimental language model I trained from scratch, running offline on ESP32-S3 with a touchscreen chat interface.
A decision tree on an STM32 spots when you drink and reminds you when you don't. No internet, no cloud.
An open ESP32-S3 camera in a hanging solar trap photographs fruit flies over 4G; OpenCV 5 counts them and a person approves the next move.
RT-Spark streams motion data over MQTT; a Flask + CNN backend classifies walking or running, shown live in an Android app.
A glowing pumpkin that spots visitors, hears their questions, describes what it sees and replies spookily, running models on Dragonwing NPU.
A chest X-ray goes in, a written report comes out, on one Jetson Orin Nano Super. Two models share 8 GB, and nothing leaves the board.
A small battery powered sensor box that straps to the casing of a domestic surface water pump
A Bluetooth bearing monitor that relates vibration to shaft speed and shows evidence for inspection using sensors I already own.
A clamp-on Bluetooth sensor that hears an aeration pump wearing out and warns before fish or septic tanks run short of oxygen.
Predicts borewell pump failures and checks water quality, on-device with the nRF54LM20 DK over Bluetooth.
A family photo server on a Pi 5: full-resolution phone backup, on-device AI search and encrypted RAID 1 storage on two USB drives.
WaterLens combines four water sensors, Arduino UNO Q and local Edge AI to recognize patterns and flag changes from a tap-water reference.
A Pi 4 and a cheap USB mic tell dialogue from explosions with an on-device voice model, then click your TV's volume over Wi-Fi or IR.
A camera-based RC car using a PilotNet-style CNN and LiteRT on Raspberry Pi to predict steering, with gamepad-controlled throttle.
A bike tag that feels the road, classifies the surface with on-device AI, and maps cycle paths that are wearing out before potholes form.
A retro-inspired math adventure where kids use camera-based finger input to solve challenges and save the Kingdom with Arduino UNO Q.
This project demonstrates cooperative robotics with multi-agent reinforcement learning (MARL) policies deployed for embedded inference.
Train LeRobot policies from real demonstrations and deploy on the Dragonwing IQ-9075 NPU for autonomous pick and place task.
Open RTL, ISA, simulator and compiler, running LFM2.5 and Qwen3 on a repurposed Kintex-7 PCIe card.
Inspired to develop this local AI agent after failing to remove a yellowjacket nest with my custom bee vacuum due to my lack of knowledge :)
In this project, we take a two-stage OCR pipeline originally built in Python and migrate it to the Arduino UNO Q using Edge Impulse.
Autonomous robotic dog leverages ROS 2 and Edge Impulse-trained YOLO-Pro model for real-time ball detection and pursuit on the Arduino Uno Q
AI-Powered Smart Tracking Camera
