Member of Technical Staff - Vision-Language Models

Experience


Skills

• Vision-language models
• Multimodal ML
• PyTorch
• Grounding
• Clinical evaluation
• Agent harnesses

What You Will Build

• Vision-language models for radiology: adapting open and frontier multimodal architectures to imaging-report data, from fine-tuning through preference optimisation on real radiologist edits
• Grounding systems that tie every generated clinical statement to image evidence - localisation, slice and region attribution, and explicit uncertainty when the evidence is weak
• Evaluation against clinical ground truth: benchmark suites built with our radiologists, regression catches across model versions, calibration surfaces, and error taxonomies that separate harmless phrasing drift from clinical error
• Agent harnesses in production: orchestration, tool integration, structured state passing, retry and fallback logic, deterministic replay, and cost / latency governors around the models you ship
• Data pipelines across DICOM studies, reports, annotations, and radiologist feedback that turn daily reporting into training and evaluation signal
• Tight loops with the computer vision team and the radiologists who define ground truth. The model is only as good as the evaluation it answers to

You Should Have

• 2+ years of hands-on experience training, adapting, or productionising multimodal or large language models - not just calling APIs from a notebook
• Strong Python and PyTorch, including custom datasets, training loops, distributed training or inference, and serious debugging instincts
• An eval-first mindset: you instinctively reach for reproducible measurement against ground truth before changing a model, a prompt, or a dataset
• Experience with at least one of: vision-language architectures (contrastive, LLaVA-style, or natively multimodal), image-conditioned generation, grounding and localisation, RLHF / preference optimisation, or clinical NLP
• Clear thinking about how a generative model fails in a clinical setting: hallucinated findings, false reassurance, laterality errors, silent regressions

Even Better If You Have

• Publications, strong research artifacts, or open-source contributions in multimodal ML or medical AI
• Experience with medical imaging data: DICOM, 3D volumes, radiology reports, or clinical annotation workflows
• Built evaluation harnesses, tracing, or trajectory tooling for LLM or agent systems running in production
• Worked with clinicians to define ground truth in a high-stakes domain

Disclaimer: This job posting has been aggregated from external source. Role details, content, and availability are subject to change. Applicants are advised to confirm the latest information directly on the company website before applying.