Final-year Computer Science undergraduate at the University of Bucharest, expected to graduate in July 2027.

I lead a project on how much Romanian grammar small language models learn from media made for children. I built a 10M-word corpus from the captions of Romanian children's videos, trained GPT-2 and GPT-BERT models on it and evaluated their grammatical agreement.

• Radu-Stefan Popa, Adina Camelia Bleotu, Sergiu Nisioi. Romanian Small Language Models Trained on Native Child-Targeted Media. NeurIPS 2026 Workshop on Linguistic Principles for Foundation Models (LP4FM), accepted poster, non-archival. Code

PQPP is a benchmark for predicting how hard a prompt is for text-to-image generation and image retrieval, and we are extending it to more languages. I built the image-captioning pipeline used for six languages (BLIP-2 and Centurio, zero-shot and with LoRA), and I run the Diffusion-DPO fine-tuning of the text-to-image models across languages.

I work as a software engineer on multithreaded C# and C++ software for eye surgery systems, where I also built an LLM assistant for bug triage. At the University of Bucharest, I am an undergraduate tutor for algorithms and data structures.

OpenReview · LinkedIn · Email