Liliana Hotsko

About

I am a Graduate Computer Science student at the University of Waterloo, supervized by Yuntian Deng. My research interests include Large Language Models, privacy‑preserving NLP, and agent systems.

Waterloo, ON, Canada · Open to research collaborations and internships.

Selected Works

Program As Weights

LLM Neural Networks

Research on representing programs as neural network weights, exploring novel approaches to program synthesis and neural architecture design.

Paper in Submission

LLM Anonymization Framework

PII masking RAG HE

A modular framework for pre‑LLM anonymization with transformer‑NER + LLM masking, encrypted vector search, and RAG evaluation. Supports English & Ukrainian, with a small UI for interactive testing.

GitHub Thesis

Hyper/LoRA vs Full‑FT (RepoBench)

Qwen2.5‑1.5B LoRA/Hypernetwork Profiling

Repository‑level QA fine‑tuning with custom masking and GPU memory profiling of LoRA, full FT, and a hypernetwork that generates LoRA weights. Focus on stable training & reproducibility.

MacTell — Open Interpreter

Swift Automation

Natural‑language control layer for macOS that executes actions (CLI, web, code, media). A unified interface to OS‑accessible tasks for productivity.

GitHub Preprint

Benchmarking: Monotonic GNNs

XAI GNNs

Extracted explanatory rules from Monotonic GNNs and analyzed how architecture & extraction parameters affect rule quality and performance.

GitHub Preprint

Contact

Email: lhotsko@uwaterloo.ca · GitHub: lilianahotsko · LinkedIn: Liliana Hotsko