An LLM agent that orchestrates a full ML pipeline: the model decides how to handle missing data, which features to engineer and which model and hyperparameters to pick, while deterministic Python nodes run the work. Tested across different models, prompting techniques and generation parameters.
Coursework reproducing LUFFY (Learning to reason Under oFF-policY guidance): a mixed-policy GRPO method that blends a model's own rollouts with off-policy reasoning traces from a stronger teacher, using regularized importance sampling so the student learns from them without collapsing into rigid imitation.
A Telegram bot that automates offline appointment booking on top of Google Calendar: clients register and cancel bookings right in the chat, while staff manage services, masters and reviews from an admin panel, with everything synced to a SQLAlchemy-backed store.