Computer Vision & Machine Learning Engineer
building production vision systems
I'm an R&D engineer who turns research ideas into scalable, production-ready AI across computer vision, multimodal and vision-language models, and large-scale learning. I focus on two hard problems: getting strong results when labeled data is scarce, and making inference pipelines fast and robust enough to deploy.
From a national geospatial AI initiative to clinical decision-support: research-grade ideas turned into systems that run.
Problem → approach → outcome. A cross-section of systems spanning satellites, slides, and microscopes.
The national Slovenian AI Factory needed analysis-ready geospatial data instead of raw, hard-to-consume archives.
Leading development of the national Slovenia Data Cube, analysis-ready, high-performance geospatial data infrastructure.
A shared, performant foundation for geospatial AI across the Slovenian AI Factory initiative.
Remote-sensing tasks tend to start from scratch, with little representation reuse across segmentation and detection.
Training foundation models for remote sensing via multi-task and self-supervised paradigms, producing reusable backbones.
Backbones power both semantic segmentation and object detection, sharing representation learning across tasks.
Optical satellite imagery alone leaves gaps; clouds and lighting limit what a single sensor sees over time.
Temporal and multi-modal fusion of Sentinel-1 radar and Sentinel-2 optical imagery.
Complementary radar and optical signals combine for more robust analysis across time.
Strong vision representations usually demand large labeled datasets many domains simply don't have.
Self-supervised representation learning combining DINO-style student-teacher training with MAE-style encoders.
Learns transferable representations without labels, the backbone strategy behind much of my label-efficient work.
Expert annotation of nuclei in pathology slides is slow, costly, and a bottleneck for downstream analysis.
Zero-shot SAM-based nuclei detection that requires no task-specific labels.
Reduces the expert annotation burden for digital-pathology workflows.
Whole-slide images are massive; extracting clinically meaningful structure at scale is hard.
DINOv2/v3/MAE embedding pipelines on multi-GPU with mixed precision, feeding cell-type clustering and survival analysis.
Turns gigapixel slides into cell-type clusters and survival signals through scalable, efficient pipelines.
Sample drift during STEM microscopy degrades image quality and traditionally needs manual correction.
A vision-based autoencoder that performs automated drift correction in STEM microscopy.
Automates a correction step that would otherwise be manual and error-prone.
The engineering discipline behind the models, reproducible from first commit to deployed artifact.
The capabilities and exact tooling I reach for, from exploratory research to reliable deployment.
Pretraining and adapting vision foundation models, with a bias toward learning from limited labels.
Bringing vision together with language and other modalities into unified models.
End-to-end annotation tooling, from pseudo-labeling and refinement to active-learning triage.
Modeling label structure and dependencies for hierarchical, multi-label, and multi-task prediction.
Scalable, reproducible training and inference pipelines built to deploy and hand off.
The day-to-day libraries and tooling behind data, modeling, and experimentation.
Where I've shipped vision systems and learned what production really demands.
Peer-reviewed work across IEEE and MDPI venues. The rigor that underpins the systems I build.
M. Stoimchev, J. Levatić, D. Kocev, S. Džeroski · vol. 62, pp. 1–16, Art. 4706416
I hold a PhD in Computer Vision and Deep Learning and work as an R&D engineer focused on turning research ideas into scalable, production-ready systems. My work addresses two recurring challenges in real-world AI: data scarcity and slow, fragile inference pipelines.
I translate deep AI research into reproducible, deployment-ready ML pipelines.
I design learning frameworks that reduce dependence on labeled data, build automated analysis workflows that minimize manual expert effort, and engineer efficient training and deployment pipelines for large-scale biological and remote-sensing datasets. Increasingly my work spans modalities, combining images with text through multimodal learning and vision-language models (VLMs), and I'm comfortable taking ideas from early research through to robust deployment in high-performance computing environments.
I enjoy working at the intersection of computer vision, multimodal and vision-language AI, medical imaging, remote sensing, and materials science, where robustness and interpretability matter as much as raw accuracy.
I'm open to Computer Vision / ML Engineer, Applied Scientist, and R&D roles, remote or relocation. If you need vision systems that move from prototype to production, let's talk.