A physical rig: an ESP32 drives a small 6 V DC motor through an L298N driver and measures its speed (slot sensor and a disc on the shaft), vibration (MPU6050 IMU) and current (ACS712).
The firmware publishes JSON telemetry every 500 ms over Wi-Fi to a password-protected Mosquitto broker. A Python consumer stores it in SQLite and a read-only FastAPI serves it.
A React and TypeScript dashboard with Recharts shows live speed, vibration and device status. A reading that is not valid is labelled as such instead of being hidden.
Four transparent threshold rules flag stale data, no speed while the motor is commanded to run, speed while it is stopped, and vibration at rest. No machine learning, and no failure prediction: there is no labelled failure data.
Built in ten phases, one subsystem at a time, with an engineering log that records every correction I had to make.
From the sensor to the screen
Every 500 ms the same chain runs end to end on my laptop and the bench rig.
ESP32Speed, vibration, current, motor state
MQTT brokerMosquitto, password, home network only
Python + SQLiteConsumer stores every window
FastAPI + rulesRead-only API, four threshold rules
DashboardReact, TypeScript, Recharts
ESP32 DevKit V1Firmware, Wi-Fi, MQTT, PWM motor control, 3 s auto-stop
L298N and 6 V motorDirection and speed by PWM, six AA cells as supply
Slot sensor and discSpeed as pulses per second, about 200 at the duty I tested
A small 1D CNN that reads a 10-second, 12-lead ECG and predicts which of five diagnostic classes apply: normal, myocardial infarction, ST/T change, conduction disturbance and hypertrophy. A record can have several at once.
Trained on PTB-XL, about 21,000 clinical ECG records, then exported to ONNX and quantized to int8 so it runs on a small CPU.
Started at the AI Night Challenge hackathon with a team of 5. The first version used a random split, which can put one patient in both train and test, so I reworked it on the official PTB-XL folds, which keep each patient in one fold.
A research and learning prototype. It is not a medical device and has not been validated clinically.
From raw ECG to edge model
One script per step, with shared settings in a single config file so input length and class names cannot drift apart.
PTB-XL10 s records at 100 Hz, five superclasses
preprocess.pyOfficial folds, normalization
1D CNNPyTorch, four conv blocks, early stopping
ONNX fp32Exported from PyTorch
ONNX int8Dynamic weight quantization
fp32 against int8
AUC on the 1,246-record held-out set in the repo, thresholds tuned on validation data. Model size is the ONNX file.
Model
Macro-AUC
NORM
MI
STTC
CD
HYP
Size
fp32
0.878
0.917
0.858
0.932
0.909
0.775
0.427 MB
int8
0.878
0.917
0.858
0.932
0.908
0.775
0.117 MB
Code and docs
Training, quantization, evaluation and inference scripts, with a demo video in the repo.
Built with a friend for the restaurant Per Voi, who asked for a QR menu where tapping a dish shows it on the table.
The 3D dishes are made with Polycam scans and Meshy.ai generation, then cleaned of mesh artifacts and compressed (WebP textures) into light GLB files.
A guest scans the QR code, browses the menu, taps “Voir en AR”, and the dish appears on their table through the phone camera. On Android, Chrome runs a real WebXR AR session with ARCore, on iPhone it opens AR Quick Look, and other devices get an interactive 3D viewer.
A full-stack web app: Next.js and TypeScript front end, Supabase for PostgreSQL, authentication, file storage and an admin area for uploading dishes, in French, English and Arabic (right-to-left).
Feedback from the restaurant: they want the 3D dishes to look more realistic, which what we are working on right now.
Try it on your phone(demo)
Scan the code, open any dish and tap “Voir en AR”. Tested on Android (Chrome) and iPhone.
A mobile-first Next.js app. Dishes are GLB models loaded with Three.js, which starts a WebXR AR session with ARCore hit-testing on Android. On iPhone it opens AR Quick Look with a USDZ file.
Resume
A one-page summary of my education, projects and skills.