Reconstruction-AI Engineer (m/f/d)

Summary: Develop machine learning models for track finding and 3D density reconstruction in an Active Muon Imaging system. Own the ML core including point-cloud and graph neural networks, differentiable simulation stacks, and uncertainty quantification.

### Summary
MuRay Tech is a deeptech spinout from DESY, Hamburg. We're hiring a Reconstruction-AI Engineer to build the machine learning models that turn detector data into tracks and 3D density images for our Active Muon Imaging system. You'll develop point-cloud and graph neural networks for track finding, physics-informed architectures for density reconstruction, and the differentiable simulation stack both depend on. Direct ownership of the ML core of the imaging pipeline, from simulated training data through to deployed reconstruction.
### The role
As Reconstruction-AI Engineer, you'll own the machine learning models that turn detector data into tracks and 3D density images. That spans track finding, density reconstruction, and the differentiable simulation tools both depend on. You will:

• Develop ML models for track reconstruction. Train on large simulated datasets, one per detector setup. Reject noise, find tracks, and estimate momentum with its error, fast enough for real-time use.
• Develop AI models to reconstruct 3D density from muon tracks. Combine a compact prior for known nominal geometry with a full-resolution field for internal defects like porosity. Build it as a physics-informed architecture, not a black box.
• Build the differentiable muon scattering simulation stack. A differentiable forward model of multiple Coulomb scattering through a 3D density field, with gradient estimators suited to the underlying stochastic transport. Validate it against GEANT4.
• Develop AI models to reconstruct 3D density directly from raw detector data. Skip explicit track reconstruction, going straight from detector hits to density. A long-term target, once high-flux data is available, built directly on the simulator and the track-based density work above.

### What you'll bring

• Point-cloud and graph neural network development. You've trained GNN or point-cloud models on labeled hit-to-track data, for noise rejection, track finding, and momentum estimation with calibrated errors.
• Large-scale synthetic training data engineering. You've built pipelines that generate, curate, and validate simulated training sets, at the scale a deep learning model needs.
• Differentiable, physics-informed reconstruction development. You've built a differentiable simulator and used it as the reconstruction engine itself. You combine a compact prior for known geometry with a full-resolution field for internal defects.
• Geant4 experience. You've validated a differentiable simulator, and the training data behind it, against Geant4 directly.
• Uncertainty quantification. You calibrate error and confidence estimates for every reconstruction. That holds whether the model is trained on labeled data or fit through optimization.
• End-to-end ML systems experience. You treat simulation, training, and deployed reconstruction as one pipeline. You've taken a model from simulated data through to production before.

### How we work

• Skunkworks approach: We work in small teams with full ownership. We keep it simple, stupid.
• Move fast: a shiny prototype means we spent too long on it. We build to learn. We polish later.
• Disagree together, commit together: we debate and challenge, regardless of title. Then we deliver the decision as one team. We win as a team. We pivot as a team.
• Courage: we always make the call. A mistake isn't a failure: it's proof we're improving. Either way, we own the outcome.
• Honesty and transparency: we share what's really happening, good or bad. That lets everyone make the right call. We stay open. We stay honest.

### What we offer

• Impact: unsolved problems, not engineering backlog. Solve them, and an inspector sees inside a sealed rocket engine in seconds instead of days.
• Pace: a model you train this week gets validated against real scan data the next. The week after, a customer sees something inside their part that no scan showed before.
• Infrastructure: access to an H200 GPU cluster in Hamburg, plus LUMI, for training and large-scale simulation. Real detector data to validate against.
• Network: development and business collaborators include DESY, German Space Agency (DLR), Siemens Energy, Marvel Fusion, Caelora.
• Ownership: an equity plan gives you a meaningful stake in the company.
• Compensation: a salary that reflects your contribution to MuRay's growth.
• Relocation: we support your move. If you're relocating from outside the EU, we help with the paperwork.

### Contact
For any additional questions, feel free to reach out to talents@muray.tech or apply directly through our application form!
We are looking forward to hearing from you.