AI Engineer & Data Scientist — computer vision, LLM and RAG systems, and the engineering around them that turns a model into something people can run.
Most of my work sits in one of three places: getting vision models to hold up on messy real-world input, building retrieval systems that cite what they claim, and wiring both into APIs and interfaces that someone other than me can use. A recurring thread is Arabic-language AI — OCR, speech, and retrieval — where off-the-shelf tooling tends to fall over.
| Project | What it does |
|---|---|
| OmniRAG | Multimodal document intelligence and RAG across PDFs, Office files, and scans — reads text and visuals, retrieves across languages, and cites the exact page behind every claim. |
| Egyptian Automotive AI | Voice-first diagnostic assistant for Egyptian Arabic. Builds a differential diagnosis, asks follow-up questions, and flags safety-critical conditions. FastAPI + React, runs locally without a paid LLM. |
| VisionPitch AI | Turns broadcast football video into structured tactical data — player and ball tracking, team assignment, and pitch calibration into real-world metres, all inferred from the video itself. |
| Blender AI Agent | Local AI-to-3D pipeline. LLM planning drives Blender through an MCP tool boundary with validation and rollback rather than arbitrary scene scripting, then exports GLB and MP4. |
| NileID OCR | Arabic OCR pipeline for Egyptian National ID cards — YOLO field detection, per-field confidence, and checksum validation of the ID number itself. |
| RadAssist AI | Multi-label chest X-ray classification with uncertainty-aware predictions, calibration, and reproducible evaluation. |
More in the repositories tab — face recognition, gesture control, video generation, and retrieval experiments.
Languages Python · TypeScript / JavaScript · SQL
ML & CV PyTorch · OpenCV · YOLO / Ultralytics · InsightFace · MediaPipe · scikit-learn
LLM & RAG LangChain · Sentence Transformers · FAISS · Qdrant · ChromaDB · MCP
Serving & UI FastAPI · Streamlit · React / Next.js · PyQt5
Practice pytest · GitHub Actions · Docker · Git LFS
I care about the parts that usually get skipped: knowing when a model is wrong, writing tests that run without a GPU or a network, documenting the limitations honestly, and leaving a repository someone else can clone and actually run.