Machine Learning Engineer

Medical Imaging AI Scale-Up | Germany | €95k – €115k + Equity | Permanent

AI Futures has been engaged by the Co-Founder & CTO of one of Germany's best-funded medical imaging AI companies to build out the machine learning team. This is a senior individual contributor hire with a clear route to lead.

The company

A venture-backed medical imaging company and CE-marked software live in radiology practices and hospitals across Europe. Their models read [CT and MRI/whole-slide pathology] in production, every day, on real patients.

They are past the question of whether the technology works. The question now is whether it works everywhere: on every scanner, in every hospital, at a volume that doubles annually.

The role

As a ML Engineer you will own models end to end - from the training pipeline to what happens when a radiologist disagrees with the output on a Tuesday morning.

This is production ML in a regulated environment. You build the model, you build the evaluation that proves it, and you own the evidence when a notified body asks how you know. The hard part is not the architecture. It is generalisation: a model that performs on your validation set and falls over on a scanner it has never seen is not a product.

You will be working alongside in-house radiologists and pathologists who review the output and tell you, in detail, when it is wrong.

What you'll do

• Build and ship segmentation and classification models on 3D volumes or gigapixel whole-slide images
• Own the evaluation framework - define what "good enough" means for a clinical claim, and prove it holds across sites, scanners and patient populations
• Work directly with in-house clinicians on annotation strategy and edge-case review
• Build the monitoring that catches performance drift after deployment, not before
• Produce the technical evidence that supports regulatory submission under EU MDR

What you'll bring

• Production ML, not research ML - you have shipped models people depend on, and you have been on the receiving end when one failed
• Python and PyTorch - essential. Experience with nnU-Net, MONAI or equivalent medical imaging frameworks a strong advantage
• Medical imaging data in the real world - DICOM that does not conform to spec, inconsistent tagging, ground truth two experts disagree on
• Evaluation rigour - you are as interested in how the model fails as in how it performs
• Comfort working with clinicians who will challenge your output directly
• EU MDR or FDA submission experience
• Whole-slide image handling at gigapixel scale, or 3D volumetric segmentation
• Foundation models applied to medical imaging

Permanent | Hybrid.