All Projects

Text-to-CAD platform with an Egyptian Building Code (EBC 2023) quality gate and RAG layer (A+ / 99% graduation project, 255 commits, live demo). Turns Arabic/English prompts into architectural floor plans exported as DXF. Features 40+ API endpoints, 30+ React components, 120+ tests, Docker deployment; estimated 85–90% faster than manual CAD drafting.

Hybrid retrieval (Qdrant dense + BM25 + exact-article lookup, RRF k=60) over all 1,149 Egyptian Civil Code articles with grounding guardrails. Compared 8 configurations in MLflow, CI RAGAS gate (faithfulness 0.90), and Locust load-tested at 49.4 RPS (P95 418 ms, 0% errors over 14,820 requests).

Config-driven end-to-end MLOps pipeline on 568K transactions: data leakage audit (excluded ID feature correlated r≈0.86 with label), stratified 5-fold CV, and metric-based quality gates (PR-AUC, precision/recall) evaluating Logistic Regression, Random Forest, and XGBoost. Features MLflow experiment tracking & model registry, FastAPI inference API with Pytest test suite, Docker containerization, and ONNX runtime export with latency benchmarking.

QLoRA fine-tune (NF4, r=16) of Qwen2.5-1.5B-Instruct on a curated Ashaar-based instruction set (10 poetic themes, 1–4 verses) using a single T4. Published on Hugging Face with an interactive streaming Gradio demo Space.

Open reference notebook covering 16 quantization sections (from-scratch affine/group-wise, GPTQ, AWQ, GGUF K-quants, bitsandbytes NF4/LLM.int8, HQQ, SmoothQuant, KV-cache, QLoRA, QAT) benchmarked on model size, Wikitext-2 perplexity, tokens/s, and peak VRAM on a free T4 GPU.

Full-stack RAG SaaS (114 commits): PDF/TXT ingestion, chunking, embeddings, pgvector/Qdrant search, switchable LLM providers (Cohere/OpenAI), JWT auth, per-user knowledge bases, Prometheus metrics, and Docker deployment.

Analyzes any GitHub repo: LLM architecture summary, interactive dependency graph (files → classes → functions), hybrid static + LLM bug detection, refactoring diffs, and RAG Q&A with citations. Built with Clean Architecture (Factory/Visitor/Strategy), Celery workers, WebSocket progress, and 18 tests.

Ubuntu CLI that scaffolds NLP/CV/ML projects, inspects datasets (CSV, Excel, JSON, Parquet) and diagnoses Python/CUDA/GPU health. (5 stars, 3 forks on GitHub).