SoilNext builds the data layer for soil biology - AI microscopy that turns a slide into numbers a farmer can trust, in production today on real farm samples. Two-person team, Dutch patent filed, real commercial traction.
The role
You'll own the computer vision that turns raw microscopy into a trustworthy number, reporting to the CTO. The pipeline already exists and works - trained detectors plus custom morphometry. The near-term work is on both concrete and sequenced roadmap and ample open-ended research.
What we're looking for
- Deep, hands-on detection / instance segmentation - you move the metrics that matter through real error analysis, not just a training loop.
- Model efficiency and distillation - a genuine feel for FLOPs and what makes CPU inference fast.
- Rigour about measurement - you anchor claims to a measured ceiling and would rather report an honest number than a flattering one.
- A track record on heavily imbalanced, long-tailed data - performance on the tail, not just the head.
- Genuine command of classical computer vision (resolution, image quality, morphometry), used where it beats deep learning.
- Strong Python and PyTorch; comfortable with Google Cloud, Docker, and CI/CD.
Bonus: video instance segmentation and tracking; microscopy or scientific-imaging background; a real interest in soil, ecology, or regenerative agriculture; ML shipped inside an early-stage startup.
You're allergic to overclaiming, pragmatic about deep-learning-versus-classical trade-offs, and you think for yourself - you use AI tools like everyone does, but you know where to overrule them.
To apply
A short note on who you are and why this fits, your CV, and anything that shows your work (GitHub, papers, a model you're proud of). If AI tools are part of how you work, tell us about cases when you overruled them. Email [email protected], subject: Computer Vision Engineer.