I build AI systems from the data layer to the interface—grounded assistants, fine-tuned models, and machine-learning workflows evaluated against real baselines.
Four projects across retrieval, language-model adaptation, classical ML, and market prediction. Each one is scoped to a concrete question.
Crawls and indexes site content, retrieves relevant passages through hybrid search, and answers with source citations. The same backend powers an embeddable chat widget, private knowledge bases, and a Telegram bot.
Fine-tuned Qwen3-VL-8B on a curated and synthetic passage-to-flashcard dataset. The adapter learned to generate fewer, more selective cards while preserving valid JSON and the source text.
Exact card-count alignment on a held-out 20-passage evaluation · 3× improvement
Built a point-in-time customer dataset and compared logistic regression, random forest, and XGBoost using rolling validation. Evaluated on an untouched later period and calibrated the selected model’s probabilities.
A modest signal on synthetic data; reported against the rare-event baseline, not as a production claim.
Explored binary classification for whether a Steam item’s price would rise or fall over 1-, 7-, and 30-day horizons.
Start with a concrete user or business question.
Design the data and evaluation before choosing the model.
Ship an interface people can actually use.