Suhaib Bin Yousouf

Leading architecture for Avyuct's Universal Vascular World Model (VWM) on remote multi-node NVIDIA Blackwell/H100 GPU clusters: reproducible training, validation, and paths to regulated deployment.

• Architecting the Universal Vascular World Model (VWM) — organ-agnostic AI foundation trained via self-supervised I-JEPA on unlabelled CTA/MRI volumes, enabling generalisation across scanners and patient populations without labelled data.
• Building distributed training pipelines on remote multi-node NVIDIA Blackwell/H100 infrastructure with FP4/FP8 precision optimisation, plus high-fidelity synthetic data generation for robust pre-training.
• Directing the full pipeline from pre-training through clinical validation, targeting sub-50-case fine-tuning for new tasks (vs. 5,000 cases today).
• Extending the VWM from brain to coronary and renal vascular beds — one foundation model replacing dozens of single-task tools.
• Coordinating with clinical and regulatory stakeholders on sandbox-style evaluation and evidence packages suitable for health authorities.

At Avyuct Inc. (Virginia), developed core AI models across Avyuct's portfolio — LVO, DMVO, and SVO programs — driving multiple U.S. FDA regulatory submissions and four US provisional patent applications on which I am a named inventor.

• Built LVO detection model on CT Angiography — core to Avyuct's FDA CADt application for acute ischemic stroke triage.
• Developed Avyuct DistalPoint (DMVO): 86% sensitivity (95% CI: 73.3–94.2%) for DMVO localization (detection + localization), 84% specificity on a 75-case multi-site dataset — actively under FDA Breakthrough Device review (Q260802).
• Built Small Vessel Occlusion (SVO) model for detection and localization on CT Angiography — Pre-Submission / 510(k) path; regulatory filing planned soon.
• Developing Hemorrhagic Stroke AI — regulatory filing planned soon.
• CT Perfusion analysis pipeline — in active development (clinical validation and regulatory path to follow).
• Designed pneumothorax quantification pipeline on chest X-ray — 86% sensitivity.
• Integrated LLM-based radiology report generation via MCP server; contributed to Avyuct's US provisional filings listed in the Patents section.

Led end-to-end computer vision solutions across a 200-camera industrial surveillance infrastructure — from conception through production deployment on edge hardware.

• Invented and deployed a Machine Guarding system covered by an Indian patent application (202431015004), filed through Tata Steel Ltd., using YOLO-based virtual barricading — first AI safety system of its kind in the deployed steelworks environment.
• Built Caster Breakout Detection on NVIDIA Jetson Xavier with GCP integration for real-time monitoring and automated alarms, reducing costly downtime.
• Developed person counting (YOLOv8 + tracking), ladle hook engagement detection, slip/trip detection, and stationary object detection — all in production.
• All systems deployed on edge hardware (Jetson Xavier, Raspberry Pi) with C++ pipelines, TensorRT, ONNX quantisation for real-time inference.