I build AI systems that turn research ideas into working products: visual recognition under limited supervision, real-time computer vision, autonomous robotics, model evaluation, and deployable ML pipelines.

• Few-shot visual learning: recognizing new categories from small reference sets without full retraining.
• Real-world computer vision: detection, segmentation, retrieval, and evaluation pipelines for messy deployment settings.
• Autonomous systems: perception and simulation for drone inspection, landing, and reporting workflows.
• Applied ML tooling: semantic matching, fraud-risk metrics, and data-driven decision support.

Goal: automate cafeteria tray recognition and pricing in a real restaurant environment.

• Built an end-to-end CV pipeline: preprocessing, detection, segmentation, few-shot retrieval, classification, and post-processing.
• Reached 86.4% accuracy on a 596-image held-out set across 32 dish classes.
• Designed a few-shot retrieval stage using DINOv2/SigLIP-style embeddings so new dishes can be registered from a small photo set without retraining.
• System is being deployed in the KFUPM campus restaurant.

Goal: reduce manual road inspection and documentation through autonomous drone reporting.

• Built a drone inspection system for road-damage detection, precision landing, and automated report generation.
• Trained custom road-damage models with 0.88 recall, 1.00 precision, and 47.8 mAP@50.
• Implemented a Dockerized SITL simulation stack for Pixhawk/ArduPilot precision landing on ArUco markers under wind and safety constraints.

Goal: help small merchants access lease-to-own equipment financing from operational and market signals.

• Built a Dockerized cloud agent for underwriting analysis and merchant financing insights.
• Combined POS-data deep dives, financial document extraction, Google Maps review scraping, market indicators, rule-based checks, and payment APIs.

• Built continuous Arabic sign-language recognition models using TCNs, Transformers, and GNN variants, achieving 15.7% WER.
• Won a tabular ML competition with 82% minority-class F1 using TabPFN, AutoGluon, custom deep models, focal loss, threshold optimization, and feature engineering.
• Built a Jetson Nano drone-follow system using YOLO, FaceNet reference-image matching, and midpoint tracking.

• Keeta intern: built a fraud-risk metric that flagged around 400 high-probability cases and shipped a semantic text-matching tool that reduced a week-long manual workflow to minutes.
• Competition wins: ByteBank Technical Hackathon, Four Principles Consulting, and KFUPM MBA Algebra Contest.
• Saudi Universities Chess Champion: 2x gold and 1x silver representing KFUPM.

AI / ML: PyTorch, TensorFlow, Transformers, scikit-learn, YOLO, DINOv2, CLIP-style embeddings, VLMs, LLMs, RAG, LangChain, TabPFN, AutoGluon, LoRA, fine-tuning, model evaluation.

Engineering: Python, Java, C/C++, JavaScript, SQL, NoSQL, FastAPI, Docker, REST APIs, Git, Linux, CI/CD, Firebase, Power BI.