Applied Research Scientist: Medical AI Consulting

Architecture design and SSL pretraining strategy for a ViT-based foundation model on 5M+ gastrointestinal video frames for endoscopic polyp detection.

Scope (NDA-safe)

• Designed SSL pretraining strategy combining masked image modelling with DINOv3-style self-distillation.
• Built data ingestion pipeline from cloud object storage to training-ready datasets under strict GDPR compliance.
• Set up experiment tracking and reproducibility workflows across multiple model architectures.
• Defined fine-tuning protocols for classification, segmentation, detection, and severity scoring tasks.
• Managed cloud GPU resources to optimize training throughput and cost-efficiency.
• Planned deployment pathways for clinical validation and edge-oriented inference environments.

Stack

• Python, PyTorch, Vision Transformers (ViT), DINOv3-style self-distillation.
• Weights & Biases for tracking and experiment management.
• S3-backed datasets and cloud GPU compute.

Engineering Challenges

• Processing high-volume, heterogeneous endoscopic video under medical data governance constraints.
• Designing multi-objective pretraining that balances feature self-distillation with masked image modelling.
• Reducing bottlenecks between storage, preprocessing, and multi-GPU training jobs.
• Keeping experiments reproducible across rapid model iterations and changing dataset versions.

Outcome

• Foundation model architecture capable of self-supervised learning on millions of unlabeled clinical frames.
• Benchmarking framework comparing performance against state-of-the-art medical foundation models.
• Reusable training pipeline scalable across future medical imaging tasks and institutions.
• Foundation for real-time, clinically relevant inference pipelines under GDPR and EHDS compliant data governance.