Computer Vision · Vision-Language Models · NLP
Informatics undergraduate and INSAIT research scholar working on vision-language models for satellite imagery and 3D scenes; previously a Quant ML intern at Man Group and an NLP research intern at Graphwise, with a deep interest in aviation safety, originally from Lyaskovets, Bulgaria.
I am an undergraduate research scholar at INSAIT and Sofia University, advised by Dr. Danda Paudel, working on remote-sensing vision-language models and spatial reasoning in 3D scenes.
I recently started working on GPU kernel programming, with a first project on efficient transformer inference. Previously, I was a Quant ML intern at Man Group and an NLP research intern at Graphwise. Earlier work covered synthetic-image detection, adversarial robustness, and interpretability.
Outside research, I am interested in aviation safety and aircraft design — especially the Airbus A350. For now I follow interesting aircraft on Flightradar24; spotting them from the ground - soon.
Sofia, Bulgaria
What I am exploring
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
Aligning 3D point-cloud and Gaussian representations for structured spatial reasoning and scene understanding.
With Runyi Yang · Advised by Dr. Danda Paudel
Where I have worked
Worked on table entity reconciliation: matching entities across tabular records and linking them to a knowledge base.
Lecturing and mentoring students preparing for the International Olympiad in Artificial Intelligence (IOAI).
Doing research at INSAIT in the area of Computer Vision — mostly working on 3D vision processing and Vision-Language models.
Conducted Quant ML research, focusing on developing and evaluating machine-learning models for systematic trading and signal forecasting.
Peer-reviewed work & preprints
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.
Things I have built
A streaming Triton kernel for exact DeBERTa-v2/v3 attention that avoids materializing the full attention matrix. On H200 benchmarks, it is up to 2.33× faster and uses up to 91% less peak memory than Hugging Face Transformers eager.
A secure, bilingual platform for machine-learning competitions, built for Bulgarian AI Olympiad selection and national contests. It runs notebook and Python submissions in hardened Docker sandboxes and provides live leaderboards, double-blind jury review, custom evaluators, and distributed workers.
A from-scratch PyTorch Transformer for Bulgarian-to-English neural machine translation, developed as coursework in Information Search and Retrieval: Application of Deep Learning at Sofia University. It combines custom BPE tokenization, masked multi-head attention, KV-cached beam search, and training optimizations, achieving 42.57 BLEU-4 with 12.26 test perplexity.
A dependency-aware C++ expression engine with polymorphic tokenization and two-stack infix evaluation.
A modular C++ audio workstation with timeline editing, file I/O, and an extensible effects architecture.
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.
Academic background
Bachelor of Informatics
Recognition
What I work on