Job
In this role, you develop vision models using PyTorch, optimize training pipelines, and evaluate models for efficient inference on edge devices while applying innovative ML techniques.
This listing has been machine translated. Show original
Job technologies
Your role in the team
- You develop and train our domain-specific vision models based on state-of-the-art architectures, using our waste image dataset for pretraining and continued pretraining.
- You design our entire ML training pipeline: from data preparation and distributed training (PyTorch FSDP/DDP, mixed precision) to model versioning.
- You build and maintain our eval suite—the central infrastructure that measures whether our models are actually getting better: linear probing, k-NN probing, few-shot detection, cross-domain generalization, anomaly detection.
- You fine-tune and distill our models for specific downstream tasks and edge hardware (sorting systems, GPU inference).
- You systematically analyze training runs, identify issues such as feature collapse or domain shift, and develop sustainable solutions instead of short-term workarounds.
- You work closely with the cloud backend team to deploy models efficiently (ONNX, TensorRT, OpenVINO).
- You actively follow research developments in computer vision and translate relevant papers into production-ready solutions.
- You think beyond the model and keep in mind how your work performs in real-world operations—for sorting facilities, customers, and the system as a whole.
Our expectations of you
Qualifications
- You have a strong command of PyTorch, including distributed training (DDP, FSDP), mixed precision (bf16/fp16), and performance optimization (torch.compile, profiling).
- You understand not only how to train a model, but also how to evaluate it. You know that a weak eval suite makes all pretraining worthless.
- You use modern AI tools (e.g., Claude, Copilot) to speed up routine coding and focus on the truly challenging research and architecture questions.
- You have a good understanding of data pipelines for large datasets (millions of images): efficient data formats, GPU augmentations, and I/O bottlenecks.
- You have expertise in inference optimization and model distillation, and ideally have already deployed models on edge hardware.
- Strong problem-solving skills, analytical thinking, and scientific rigor—you work based on hypotheses rather than trial and error.
- Confident working with cloud GPU infrastructure (AWS, Azure, GCP, or on-premises H100/A100 clusters).
- Fluent German and good English skills are required.
Experience
- You have several years of experience developing and training computer vision models, ideally using modern Vision Transformer architectures and self-supervised learning methods.
- You have experience with modern ML tooling stacks for configuration management, experiment tracking, data versioning, and backbone libraries.
- Experience with common detection/segmentation and anomaly detection frameworks is an advantage.
- Ideally, you have research experience of your own (papers, open-source contributions, conference talks) or a PhD—neither is required, but both are a plus.
What we offer
- Building from the ground up — developing our own foundation model strategy without legacy systems or technical debt.
- Access to a unique, growing database of real-world sorting facilities—a strategic advantage that no university and hardly any competitor can match.
- Use of current frameworks and a state-of-the-art tech stack (PyTorch 2.x, FSDP, Hydra, W&B, DVC, timm).
- Substantial compute resources for pretraining runs—we know that serious foundation model training isn’t a hobby project.
- Close collaboration with research partners (including THWS Würzburg as part of Green-INNO) and the opportunity to publish your own research findings.
- Work in a dynamic, interdisciplinary start-up team with plenty of responsibility and scope to shape things from day one.
- Short decision-making processes and direct communication.
- Technology with purpose: You work on the biggest levers of our time—AI, recycling, and the circular economy.
Topics you will work on
Job locations
About your employer
WeSort.AI GmbH, based in Würzburg, utilizes innovative technologies such as AI, cameras, and X-ray technology to accurately identify and sort waste. Their systems contribute to safety by detecting hazardous materials and support the recovery of essential raw materials for the circular economy. Clients include major waste management corporations and municipal disposal services.
- Startup
- Onsite
- Industry, Production
More jobs
These jobs might also interest you
You actively shape the technical development of our software products and influence their long-term architecture. You develop our core applications primarily using C# as the central programming language and technological foundation. You design, develop,...
Hello, let's say WE! We - the HUK-COBURG Insurance Group - are among the top 10 in Germany. Founded over 90 years ago, we are now the leading insurer for private households with over 13 million customers and the largest German auto insurer. Our...
sino AG, based in downtown Düsseldorf, has been an innovative financial service provider for over 25 years. We specialize in the execution of securities transactions for the most active and demanding traders in Germany. With currently 26 employees, we...
