A Machine Learning Engineer with a Computer Engineering foundation, specializing in Computer Vision, NLP, and Large Language Models. I build detection, tracking and multi-agent systems — and I like reproducing research papers until the numbers beat the paper.
I'm an AI & Machine Learning Engineer with a Computer Engineering foundation, specializing in Computer Vision, Natural Language Processing, and Large Language Models. My work spans real-time detection and tracking, medical imaging, and multi-agent LLM pipelines — built with Python, PyTorch, LangGraph, LangChain, SQL, Docker, and AWS.
I have a strong track record of independent research — reproducing landmark papers (ECCV, CVPR) from scratch, running ablations against multiple baselines, and pushing accuracy past the reported numbers — alongside cross-functional delivery of production-ready AI systems.
A mix of production systems and from-scratch paper reproductions — each tagged with its live confidence, i.e. the reported result.
Real-time, sensor-fused counter-UAS platform: dual YOLOv8 detection across RGB + thermal feeds fused via calibrated similarity-transform alignment, an NMSE-based threat-prioritization engine, and full PID visual servoing with an optional online Q-learning tuner. Ranked Top 5 of 320 teams at ITC Egypt ADC 2026.
Multi-agent system with specialized agents for question generation, summarization, and feedback-driven refinement, plus a RAG pipeline over course materials with async FastAPI microservices, Whisper transcription, and PyMuPDF parsing.
Re-implemented a Hierarchical Relational Network on 4,830 frames from 55 videos via a custom Graph Relational Layer — 91.55% test accuracy, pushed to 92.00% by adding a Graph Attention (GAT) mechanism with test-time augmentation.
Hierarchical Temporal Deep Model with a two-stage LSTM to capture multi-person temporal dynamics — 93% accuracy on benchmark data, with ablations isolating each component against 8 baseline models.
Three progressively stronger baselines — U-Net → improved U-Net → DeepLabV3+ with ASPP — on CamVid across 12 classes, taking mean IoU from 0.34 to 0.59 via Dice+CrossEntropy loss and mixed-precision training.
Failure-forecasting system across 10+ sensors predicting equipment failures 24 hours in advance — 84.7% recall, 96% accuracy, tuned to minimize false alerts and deployed via containerized MLOps workflows.
MultiCheXNet: a multi-task DenseNet-121 performing simultaneous classification, detection, and segmentation through a shared encoder, trained across datasets with mismatched label schemas via a custom dataloader.
Tackled extreme class imbalance on 284,807 transactions with a voting-classifier ensemble and a focal-loss-enhanced PyTorch network — 86% F1, 85% PR-AUC, outperforming standard classifiers on rare-event detection.
6th International Telecom Conference (ITC Egypt), hosted by the Egyptian Air Defense College. Represented Team Skyhawks, building and presenting the Multi-Layer Anti-Drone Defense System as the team's competition entry — ranked in the Top 5 among 320 competing teams, advancing past an initial shortlist of 45 qualified teams.
ITC Egypt ADC · 2026