Efficient Machine Learning · Vision-Language Models · 3D Scene Processing
I am an Informatics student at Sofia University and an undergraduate research scholar at INSAIT, advised by Dr. Dimitar I. Dimitrov on efficient machine learning, and by Dr. Danda Paudel on vision-language models for remote sensing and 3D scene processing. Previously, I was a Quant ML intern at Man Group and an NLP research intern at Graphwise.
Outside research, I am interested in aviation safety and aircraft design, especially the Airbus A350. In my free time, I collect and build scale models of airliners and follow interesting aircraft on Flightradar24.
Sofia, Bulgaria
Accelerating relative-position attention with exact, fused Triton kernels, using offset-aware tiling, cross-tile reuse, and an atomic-free backward pass for faster, more memory-efficient training and inference.
With Ivo Petrov · Advised by Dr. Dimitar I. Dimitrov
Adapting vision-language models to satellite imagery using scalable supervision derived from OpenStreetMap.
With Mario Markov, Stefan Maria Ailuro, and Mohammad Mahdi · Advised by Dr. Danda Paudel
Data-efficient structured 3D scene understanding, building on SpatialLM with a Chorus-based 3D Gaussian encoder to lower training-data requirements.
With Runyi Yang · Advised by Dr. Danda Paudel
Bulgarian National Olympiad in AI
Member of the national committee organizing the Bulgarian National Olympiad in AI: writing competition tasks and running contests on LavBench.
Worked on table entity reconciliation: matching entities across tabular records and linking them to a knowledge base.
Conducted Quant ML research, focusing on developing and evaluating machine-learning models for systematic trading and signal forecasting.
An exact FlashAttention-style tiled algorithm that brings additive relative-position attention (Shaw, Transformer-XL, DeBERTa, T5) into fused Triton kernels. It projects only reachable offsets, reuses relation entries across constant-offset tiles, and accumulates gradients by offset without atomics: up to 6.4× faster forward and 21× faster backward at 8K tokens.
A self-contained adaptation framework that uses OpenStreetMap data to generate geographic supervision without manual labels or a stronger external teacher.
A balanced 200K-image benchmark spanning four content types and multiple open and proprietary generators, with strong contrastive-learning baselines.
An empirical study of how adversarial training changes learned representations and makes computer-vision model decisions more interpretable.
A streaming Triton kernel for exact DeBERTa-v2/v3 attention that avoids materializing the full attention matrix. Benchmarks show up to 5.18× faster forward and 2.15× faster forward+backward than Hugging Face Transformers eager. Integrated into GLiNER2 (fastino-ai) as an optional attention backend, released in v2.1.0. The arXiv paper Fast and Exact Relative-Position Attention via Structure-Aware Kernels (under review) is a continuation of this project.
A secure, bilingual platform for machine-learning competitions; the official evaluation system of the Bulgarian National Olympiad in AI. It runs notebook and Python submissions in hardened Docker sandboxes and provides live leaderboards, double-blind jury review, custom evaluators, and distributed workers.
A 144 GB, 200K-image dataset for synthetic-image detection, balanced across photos, paintings, faces, and uncategorized imagery. It supports both real-versus-synthetic classification and generator identification across open and proprietary image generators.
A PyTorch implementation of gradient caching for scaling computer-vision contrastive learning beyond GPU memory limits. It reproduces large-batch gradients from smaller chunks and supports SimCLR, SupCon, SelfCon, mixed precision, and distributed training.
PyTorch training and evaluation tools for studying how adversarial robustness affects model interpretability, with SHAP, Integrated Gradients, representation inversion, and class-specific visualization workflows.
Bachelor of Informatics