Production-grade ML systems. Real results.
Multi-provider LLM chatbot supporting 10+ models (Cohere, OpenAI, Groq) switchable from the UI. Supports 5 prompting strategies: Zero-Shot, Few-Shot, Chain-of-Thought, Memory Chain, and Structured JSON. Streams responses via SSE with per-session memory, 6 AI personas, and real-time cost and latency display. Deployed on Railway with FastAPI.
Serverless AI assistant deployed on Cloudflare Workers, representing Mohamed professionally 24/7 at zero infrastructure cost. Implements a three-model fallback chain (LLaMA 3.3 70B → 3.1 70B → 3.1 8B) across independent Groq token pools. Uses keyword-based CV extraction to send only relevant sections per query. Supports bilingual Arabic/English input with Arabizi detection.
Fine-tuned Qwen 2.5 7B VLM with LoRA rank-32 on 35,000 retinal images. Multi-agent pipeline (DenseNet + EfficientNet + RBF-SVM) achieved 240% accuracy boost (20% → 68%). 4-bit QLoRA cut VRAM 17 GB → 7 GB with no accuracy loss. RAG with FAISS reduced hallucinations 18% → 5%.
RAG pipeline indexing 20+ medical textbooks (5,000+ pages) with FAISS. 0.3 s retrieval latency, 0.89 recall@5. Grounded VLM outputs lifted retinopathy accuracy to 68%, edema to 94%. Improved BLEU 0.17 → 0.26 and cut hallucination rate from 18% to 5%.
Flask app delivering side-by-side retinal fundus analysis from ResNet50 and a Vision Transformer (ViT). Dual-output classification: 5-level retinopathy severity (No DR → Proliferative DR) and 3-level macular edema risk. Trained on the IDRiD dataset with PyTorch and the timm library.
Custom CNN achieving 95% accuracy on 3,000 brain MRI scans across 4 tumour types (glioma, meningioma, pituitary, healthy). 400% dataset expansion via augmentation. Dockerised Streamlit app with real-time inference and Class Activation Map visualisation.
Custom hybrid encoder (GoogleNet + ResNet-18 fusion) with triplet loss trained on VGGFace2 (9,000+ identities). 512-dim L2-normalised embeddings, 95%+ validation accuracy, 50 ms GPU inference. Desktop Tkinter app with SQLite identity database (1,000+ profiles) and Streamlit web demo.
Graduation project benchmarking 21 deep learning architectures (LSTM, GRU, BiLSTM, CNN-LSTM, Transformer, TFT) for solar and wind forecasting. 98% accuracy, less than 15% MAPE, 48-hour horizon. ITAC + ASRT funded. IEEE national recognition award.