Mohamed

Aziz

Sakiss.

Building an agent that parses a resume and target job descriptions, then applies to matching roles on the user's behalf — cutting manual job-hunting time. Building a fully spoken mock-interview agent using OpenAI Whisper for listening and TTS for responding, making it feel like talking to an actual interviewer. Orchestrating multi-agent application logic in LangGraph behind a FastAPI backend serving the dashboard.

Real-Time Vision & ML Pipeline: Engineered ErgoFit to monitor live ergonomics, combining MediaPipe pose landmark extraction with a Random Forest classifier to achieve 98% posture detection accuracy. Hybrid RAG & Hallucination Prevention: Developed "ErgoCoach," an LLM assistant that fuses structured PostgreSQL data (user error patterns) with certified ergonomic PDFs to generate medically grounded advice while strictly preventing model hallucinations. Autonomous Reporting Agent: Designed a scheduled AI agent that analyzes temporal database metrics to generate and dispatch personalized weekly reports, offering actionable insights for continuous posture improvement. Architecture & Live Deployment: Built a hybrid backend (Node.js/FastAPI), containerized microservices using Docker, and automated CI/CD pipelines via GitHub Actions to launch the platform live at ergofit.live.

Built and deployed RAG chatbots (LangChain, Gemini/OpenAI) that cut support response time and manual ticket load. Built data pipelines connecting REST APIs to backend systems for real-time processing. Documented model performance and helped streamline the team's Docker/CI-CD deployment process.

Real-time perception system using YOLOv11m trained on BDD100K. Integrated SAHI to detect small distant objects with 84% Precision.

Autonomous agent bridging WooCommerce and Dolibarr ERP. Features 'Tool Use' capabilities to perform real-time SQL inventory checks.

Processed 1 Million+ transactions using PySpark. Solved class imbalance (8.7% fraud) using custom class weighting, achieving AUC > 0.99.

Developed a secure medical AI application that automates the classification of Brain Tumors (Glioma, Meningioma, Pituitary). The architecture utilizes a Parallel Ensemble Model, fusing features from VGG16 and ResNet50 backbones with a custom Channel Attention (SE) block for high diagnostic accuracy. Engineered a decentralized ledger (NeuroChain) where the model's prediction and the hash of the MRI image are instantly anchored on a blockchain to prevent data tampering. Features an end-to-end Streamlit workflow for real-time image upload, instant inference, and automatic blockchain transaction generation.

Embedded system deployed on Raspberry Pi 4. Uses a custom CNN (96% accuracy) combined with HSV filtering to eliminate false positives.