I'm a final-year student at TUES Sofia and an independent ML researcher. My work sits at the intersection of self-supervised learning, efficient / embedded AI, and representation learning for speech, vision, and world models. The recurring theme across everything I do: make small models learn well from little supervision.
- 🔭 Right now: self-supervised speech representation learning and compact-model pretraining
- 🎮 Genuinely hyped about: reinforcement learning and JEPA-style world models for planning and control
- 🌱 Digging into: value-shaped representations and action-effect encoding
- ⚙️ Comfort zone: PyTorch, CUDA, and multi-GPU SLURM/HPC training, plus shipping models to constrained hardware
- 🎓 Working toward graduate research abroad
- 🌍 Bulgarian and English
- ⚡ Fun fact: I like models with fewer parameters than my phone has contacts
- 📫 Reach me:vagrivas08@gmail.com
Self-supervised learning | Edge & efficient ML |
World models | Medical & vision AI |
- 🧮 EPU (Exponential Partial Unit) — a novel activation function, published in MDPI Applied Sciences
- 👁️ Edge vision — sub-million-parameter object detectors built for on-device inference
- 🌐 World models & SSL — ongoing research on latent-prediction objectives and value-shaped action representations
- 📄 Publications and workshop submissions across ML venues, with more in the pipeline
Languages
ML / Data
Infra & tools
"The best model is the one small enough to actually ship."
