Aboubakr Ketoun

Mobile React app with KNN-based recommendations to retrieve best-matching products in search, PayPal payments, and a Spring Boot REST backend. KNN model deployed via Hugging Face and Docker; sessions, wishlist, and cart via React Context. Added Azure ML & Azure AI Foundry plus OpenAI SDK to build a RAG chatbot.

Next.js front-end with Framer Motion; integrated Jira and Slack APIs, Stripe and PayPal payments, Google Calendar, SMTP mailing with SendGrid (Twilio) for automated confirmations/notifications, and database schema design. Deployed/hosted production stack with an Express.js backend and Supabase cloud database.

PoseAnalysis3D is a comprehensive 3D human pose estimation and body measurement system that processes standing-person images to generate detailed anthropometric data. The pipeline uses MediaPipe for accurate 2D skeleton keypoint detection (17 COCO-format joints), converts them to 3D coordinates using anatomical depth estimation, and creates realistic human avatars with colored body parts. It leverages VirtualMarker (Apache-2.0) and VirtualPose (MIT) repositories for advanced pose processing, combined with trimesh for 3D mesh generation, OpenCV for image processing, and NumPy/Matplotlib for visualization. The system outputs complete body measurements including height, shoulder width, limb lengths, circumferences (neck/chest/waist/hip), surface area, and volume—all saved as 3D meshes (.OBJ), keypoint data (.JSON), and formatted measurement reports stored in Google Drive.

Deep learning system that classifies chest X-ray images as Normal vs Pneumonia using a TensorFlow/Keras CNN. Includes rigorous data validation (PIL), preprocessing (grayscale, 128×128, rescaling), and training-time augmentation (rotation, shifts, shear, zoom, flips). Architecture: 3 Conv2D+MaxPool blocks → Flatten → Dense(128, L2=0.001) → Dropout(0.5) → Softmax(2). Trained with Adam, categorical crossentropy, accuracy metric, EarlyStopping and ModelCheckpoint callbacks. Exported to H5 for a Streamlit web app and to TFLite for a Flutter mobile app.

LSTM-based stock price prediction app with a Streamlit dashboard. Downloads Yahoo Finance data, applies MinMax scaling and windowing, uses a 3-layer LSTM with 20% dropout, and forecasts 1–30 days ahead via sequential predictions. Evaluates RMSE/MAE/MAPE and renders charts for history, test predictions, future forecast, and training loss. Supports configurable lookback, horizon, epochs, and batch size, with caching and real-time training progress.

Full-stack CRM platform for business operations management. Features include product inventory with stock alerts, supplier and client management, quote generation with draft/negotiation workflows, invoice creation with PDF export, role-based access control (Admin/Manager/Sales), revenue vs. expenses dashboards, and comprehensive sales/inventory reports. Built with Next.js 15 and TypeScript, Clerk SSO for authentication, TiDB Serverless (MySQL-compatible) for cloud database, Recharts for analytics visualization, and React-PDF for document generation. Deployed on Vercel with Genius domain and DNS management.