MSc Artificial Intelligence student at the University of Surrey (77% average, Distinction level; completing Feb 2027), with an IEEE publication in medical image classification. Available for graduate and junior ML/AI engineering roles from February 2027.
I build models that are accurate, efficient and shippable: real-time computer vision deployed with C++ and TensorRT, self-supervised and compressed models for industrial sensing, LLM agents and retrieval (RAG) tested on held-out data, LLM and NLP fine-tuning, and inference services with tests, CI/CD and monitoring.
🗣️ Arabic (native) · English (fluent)
Real-time drone video tracking: YOLO11s served by a C++ TensorRT engine, with BoT-SORT tracking.
| Project | What it shows | Highlights |
|---|---|---|
| Real-Time Drone Detection and Tracking with TensorRT | Small-object detection and multi-object tracking in drone video, deployed with a custom C++/CUDA TensorRT engine | 3.9 ms per frame (257 FPS) · +49% mAP from 1280 px training · tracking score (HOTA) 41.0 → 53.7 with 89% fewer ID switches on held-out data |
| Maintenance Copilot: Agentic RAG, Evaluated | Tool-calling LLM agent over 5,244 pages of USAF maintenance manuals, with every experiment written down before it ran and a final test on 170 held-out questions | +7.7 points correct answers over retrieve-once RAG (CI excludes zero) · faithfulness trade-off measured and reported · self-hosted gpt-oss-20b (vLLM) on par with the API model |
| RotorMamba / Lightweight PC-STE | Self-supervised Mamba encoder for machine fault diagnosis, compressed for deployment (MSc dissertation) | 95.5% Macro-F1 across 4 datasets · 1.9× faster on CPU · 1.53 MiB after INT8 quantisation |
| Review Classifier Service | Production ML: a RoBERTa API on AWS, end to end | FastAPI · Docker · Terraform · CI/CD · Prometheus/Grafana · +79% throughput with dynamic batching |
| Cross-Variety Sentiment and Sarcasm Classification | Transformer and LLM fine-tuning across English varieties (group project) | RoBERTa 0.90 Macro-F1 · Gemma-2-2B 4-bit LoRA · Gradio demo |
| Few-Shot Prompt Learning for CLIP (CoOp) | Few-shot vision-language adaptation (my component of a group project) | 64.9% → 83.5% mean accuracy over 8 datasets · ensembles & Model Soups |
📄 Publication: A Two-Stage Hierarchical Skin Lesion Classification Using Transfer Learning, IEEE ICIT 2025 (2nd author)
- ML: Python · PyTorch · TensorFlow/Keras · Hugging Face Transformers · PEFT (LoRA/QLoRA) · scikit-learn
- LLMs and RAG: Anthropic API (tool use, structured outputs) · hybrid retrieval (BM25, embeddings, cross-encoder reranking) · LLM-as-judge with human calibration · vLLM
- Computer vision: YOLO · multi-object tracking (BoT-SORT, ByteTrack) · OpenCV · CLIP · ViT
- Efficiency: C++17 · CUDA · TensorRT (FP16/INT8) · ONNX · knowledge distillation · INT8 quantisation · latency profiling · dynamic batching
- Deployment: FastAPI · Docker · PostgreSQL · AWS (EC2, ECR, CloudWatch) · Terraform · GitHub Actions · Prometheus · Grafana
📫 LinkedIn · Google Scholar · mohammmed.abdalaziz.24@gmail.com

