Efficient computer vision on limited hardware - maximising performance under tight compute budgets.
I design and train neural networks for computer vision, with a focus on semantic segmentation and multi‑label classification.
My strength is solving real‑world bottlenecks: fitting large models into limited GPU memory, cleaning noisy datasets, and pushing architectures to their practical limits.
- 🎓 Currently pursuing a degree in Applied Mathematics and Artificial Intelligence.
- 🔬 Currently exploring domain‑adaptive segmentation and efficient multi‑task learning.
- 📄 Scientific paper on domain‑specific segmentation challenges (in preparation).
- 🧪 Ongoing R&D: squeezing state‑of‑the‑art quality out of consumer‑grade hardware.
ConvNeXt‑CLF‑75 - food ingredient classifier (75 classes) built on ConvNeXt‑Tiny with CBAM attention and GeM pooling.
Trained on a heavily cleaned subset of MM‑Food‑100K (~88,600 images) for 75 classes with severe class imbalance (smallest classes contain only ~100 images). Despite this, the model achieves strong results.
| Metric | Value |
|---|---|
| Macro F1 | 0.6546 |
| mAP | 0.7142 |
| Input size | 640×640 |
| Training hardware | Single RTX 2060 (6 GB) |
The model is optimised to serve as a semantic guide for a downstream segmentation network.
But can be used completely independently as a food classifier
Custom segmentation model (100+ classes, single RTX 2060 6 GB)
Designing a memory‑efficient architecture from scratch, integrating deformable convolutions, advanced attention blocks, and deep supervision - all while keeping peak VRAM below 5.5 GB. Early results are promising; full open‑source release is planned (repo coming soon).
Computer Vision Core
PyTorch timm Albumentations OpenCV HuggingFace - used daily for model prototyping and training.
Architectural Design
Custom architectures based on CNNs, U‑Net, ConvNeXt, ResNet, deformable convolutions, ASPP, deep supervision, and more.
Memory & Performance Optimisation
Reduced peak GPU memory by over 40 % on a 104‑class segmentation model through gradient checkpointing, mixed precision, and custom memory‑efficient attention blocks. Used PyTorch profiling and torch.compile.
Data Preparation
Cleaned and consolidated ~4,000 raw fine‑grained labels into 75 balanced classes with manual merging and automated filtering. Full step‑by‑step documentation.
Deployment & MLOps
FastAPI Docker Redis PostgreSQL Git - building lightweight inference APIs and reproducible pipelines.
I’m actively looking for a full‑time Computer Vision / Deep Learning position.
- 🌍 Remote or relocation (ready to relocate worldwide, visa sponsorship required).
- 🕒 Currently in the GMT+3 timezone.
- 🇬🇧 English C1.
- 💬 Always open to discussions about exciting projects - fastest way to reach me is via Telegram.
