Younas K.

I build deep learning systems that work under real-world constraints, not demos. I have a Ph.D. in Machine Learning and 10+ years of experience across ML, computer vision, and applied AI; I also have Q1 publications and serve as a reviewer at A*/A AI venues. I take on problems where the obvious approach fails: long-tail semantic segmentation, weak/pseudo-label pipelines, frozen-backbone (DINO/ViT) + DPT architectures, and training under hard VRAM/RAM limits. R&D Engineering - End-to-end CV pipelines: semantic/instance segmentation, detection, classification (PyTorch, OpenCV) - Foundation-model adaptation: DINO/ViT backbones, SAM/SAM3 prompting, LoRA, DPT heads - Weakly-supervised & pseudo-label systems: CAM seeding, confidence-gated verification, ignore-index/ternary labels, self-training - Training engineering: loss design (Focal-Tversky, weighted CE), class-imbalance handling, multi-GPU pipelines, checkpoint/resume/early-stopping, OOM-safe execution - Diagnosis-first debugging: root-cause isolation before patching, one-variable-per-run experimental discipline Applied ML - Model design, training, optimization, and evaluation on messy real-world data - Reproducible experiments, versioned caches, documented and auditable code - Domain-specific systems (industrial inspection, standards-driven labeling) Research Delivery (on request) - Publication-ready manuscripts, methodology design, and result interpretation - Systematic literature reviews restricted to top venues, e.g., AIR - Figures, tables, and quantitative visual analytics Stack: Python, PyTorch, OpenCV, DINO/ViT, SAM/SAM3, DPT, Hugging Face, conda, Linux, multi-GPU CUDA. I deliver working systems with clean, documented code and hold strict timelines. If you have a hard CV/ML problem, especially one where off-the-shelf models underperform on your data, send the details, and I'll respond with a concrete technical approach.