About us

Sortrace builds camera-based perception for waste collection vehicles. We turn video from collection operations into insights that help municipalities and waste operators understand sorting quality, contamination, and safety-relevant items. The domain is visually messy, policy- driven, and deployed in the real world, not a clean lab benchmark.

The role

We want a senior applied computer vision specialist who can help us define and ship the next generation of our perception stack. We are not prescriptive about model family or headline architecture: we expect to explore several directions and pick what actually meets our accuracy, latency, cost, and maintainability constraints. The role can be structured as a f ull-time hire (lead-level) or a multi-month advisory / contracting engagement, depending on the person and timing.

What you will work on

• Architecture and roadmap: compare credible options for detection, classification, and segmentation under our data and deployment constraints, including how each path fits what we already operate in production.

• From research to production: turn experiments into something we can operate, monitor, and improve over time.

• Evaluation: help define how we measure quality so improvements are defensible, not just leaderboard scores.

• Deployment reality: whatever we choose must survive embedded-class hardware on vehicles, variable lighting and motion, and multi-camera setups. You should be comfortable reasoning about throughput, memory, and reliability, not only offline training metrics.

• Uncertainty and human-in-the-loop: where the product needs calibrated confidence, review workflows, or active learning, you help make that explicit in the system design.

Must-have experience

• Shipped computer vision in production (not only papers or Kaggle); you know what breaks when models leave the notebook.

• Hands-on strength across training, evaluation, and deployment tooling; you can own an end-to-end slice without waiting for a separate "MLOps person."

• Breadth across model families: real project experience with both convolutional and transformer-based vision approaches; you are not married to one camp.

• Architectural judgment: a deep understanding of the trade-offs, failure modes, and limitations of different models and architectures, and the ability to argue for one over another given concrete constraints.

• Resource-aware ML: meaningful experience getting models to run well under tight compute budgets (embedded GPUs, mobile, or other constrained targets, not only datacenter batch jobs).

• Sound experimental judgment: you can propose hypotheses, define success criteria, and kill dead ends quickly.

• Stack fluency: comfortable in Python and the PyTorch ecosystem. Specific tooling within that is something we expect to revisit together.

Nice to have

• Hands-on experience deploying on edge platforms.

• Domains with clutter, occlusion, and long-tail objects (e.g. waste, logistics, agriculture, robotics).

• Fleet or field deployment: versioning, staged rollouts, monitoring, safe rollback.

• European privacy and data-residency awareness in product design.

• Public contributions or maintainership in widely used CV / ML codebases.

Engagement

• Location: Remote within Europe is fine; Nordic time zones are easiest for day-to-day collaboration.

• Shape: Open to f ull-time senior/lead hire or a scoped advisory or contract. A short paid pilot before a longer commitment is possible if both sides want it.

What we offer

• Equity: Meaningful ownership - you share in the upside as we grow.

• Real-world impact: The work ends up on vehicles in the field, not in a benchmark. What you build affects how customers handle waste sorting at scale.

• Direct collaboration: You work with the founders and a small. No layers.

• Scope to shape: We are early enough that your architectural decisions will define the stack for years.

• Inclusive by default: We hire for competence and curiosity.

Why Sortrace

Waste collection is one of the last frontiers of urban infrastructure that has not been touched by modern technology. Waste trucks go out every day, collect waste, and come back. Almost nothing about what actually happened is captured, understood, or acted on.

We are changing that. Not with dashboards or reports, but with a perception layer that makes the invisible visible; what is in the waste stream, how sorting quality changes over time, where the system is breaking down. The data has always existed. No one has been able to read it.

This is early-stage, hands-on work with real deployment and a problem worth solving. If that combination appeals to you, we would like to talk.