I move ideas from peer-reviewed research into systems that run in production. Six years split evenly between two questions: how do you make a model provably better, and how do you make it survive contact with the real world.
I'm a Computer Vision Researcher and AI Engineer working at the Hong Kong Industrial AI & Robotics Centre (FLAIR), where I build transformer- and generative-AI-driven inspection systems and lead multimodal medical imaging research.
Outside the lab, I design and ship AI products end-to-end: a video-surveillance platform, an LLM-driven health app, and a real-time computer-vision system for boxing analytics.
My path runs through autonomous-vehicle perception at Dutch Autonomous Mobility, federated-learning healthcare research, and technical writing on generative AI systems — different rooms, same instinct: define the problem precisely, then build the smallest thing that actually solves it.
Peer-reviewed methodology and grant-funded research spanning industrial inspection and healthcare AI.
Chronic-disease patients need continuous, personalized monitoring without compromising data privacy.
Federated learning-based healthcare analytics using non-invasive biosignals, with privacy-preserving pipelines and LangChain-based LLM clinical guidance.
Predictive recovery-analytics models enabling intelligent intervention recommendations and reduced readmission risk.
Hazardous industrial environments require real-time, interpretable safety monitoring.
Thermal imaging, vision transformers, and temporal anomaly detection, with generative AI modules producing interpretable anomaly descriptions.
Enhanced situational awareness and regulatory compliance via proactive, predictive-maintenance-driven interventions.
Farooq, U. et al. "A Hybrid Deep Learning and Large Language Model Architecture for Automated Medical Image Segmentation: From Detection to Pre-Diagnosis."
Farooq, U. et al. "RT-DefectNet: An Intelligent Real-Time Framework for Surface Anomaly Detection in Industrial Systems"
Independent products taken from problem definition to a running system — no team, no handoff.
Manual video monitoring causes alert fatigue and missed incidents from high false-positive rates.
A computer-vision SaaS platform that distinguishes genuine threats from routine motion in real time, replacing manual review with automated, evidence-backed alerting.
Chronic-condition patients get isolated data snapshots between visits, with no ongoing intelligent interpretation.
Continuous biosignal tracking paired with an LLM-driven reasoning layer that turns raw readings into personalized, explainable guidance.
Boxing scoring is manual and judge-dependent, with no objective, real-time record of performance.
A live CV system fusing YOLOv8 punch detection with MediaPipe Pose over a multi-threaded RTSP pipeline, broadcasting scores via WebSocket.
PyTorch · TensorFlow · Keras · Vision Transformers · CNNs · Transfer Learning
OpenCV · YOLOv8 · MediaPipe · Segmentation · Multi-Object Tracking · SLAM · Sensor Fusion
LangChain · Prompt Engineering · RAG · Multi-Agent Orchestration · Clinical Decision Support
Docker · MLflow · Git · CI/CD · REST APIs · FastAPI · WebSocket · Edge Deployment
Python · C++ · R · MATLAB/Octave · LaTeX · Bash
Pandas · NumPy · Tableau · Power BI · Linux · CUDA · NVIDIA GPU Optimization
Open to research collaborations and senior AI / computer vision engineering roles.