Iuliia Vitiugova

Mechanistic Interpretability of Open-Source LLMs InterpretabilityLLMs

Reproduction and extension of circuit-level interpretability methods on Qwen3-4B-Instruct. Identifies interpretable feature interactions across model layers using transcoders (sparse autoencoders on MLP activations), attribution graphs, and causal interventions — ablation, activation patching, and feature steering. Introduces a formal typology of LLM computation types and a pipeline for distinguishing causal from correlational connections. Part of MPhil research at DAMTP, University of Cambridge.

Mechanistic Interpretability Dashboard First DemoVue.jsWork in Progress

Interactive web application for exploring mechanistic interpretability research results. Built with Vue.js and Vite, deployed on GitHub Pages. Makes dense interpretability findings navigable for collaborators and reviewers without requiring direct code access. This is an early first demo — the dashboard is still under active development.

Normalising Flows — Deep Learning Coursework Generative ModelsPyTorch

Implementation and analysis of a RealNVP-style normalising flow for density estimation, using affine coupling layers with alternating binary masks, trained with cosine-annealed Adam and gradient clipping. Includes flow surgery, FLOP counting, and ablation studies on depth and batch size.

Embedding Space Geometry via Triplet Loss & Bayesian Optimisation Metric LearningPyTorch

Study of how hyperparameter choices shape the geometric structure of neural network embedding spaces. Implements metric learning with batch-hard triplet loss mining and compares it against cross-entropy baselines. Bayesian hyperparameter search (Optuna TPE) is used to efficiently explore mixed architecture, learning-rate, and regularisation spaces, demonstrating improved generalisation and zero-shot transfer compared to standard classification objectives.

Electronic Nose: Calibration & Feature Analysis for Gas Sensing Neural NetworksXAI

Machine learning models for gas concentration prediction using semiconductor sensor arrays. Research covered two complementary directions:

• Feature Analysis: Compared interpretability methods — Weight Analysis, Deep Taylor Decomposition, permutation importance, and fixed-value masking — to identify which sensor measurements drive predictions.
• Neural Network Calibration: Adapted temperature scaling, vector scaling, and matrix scaling (Guo et al. 2017) to sensor drift conditions, improving probability reliability under domain shift.

Information Retrieval for COVID-19 Scientific Literature NLPIR

Systematic benchmarking of classical and neural information retrieval techniques on the TREC-COVID corpus. Compared TF-IDF, BM25, GloVe, Sentence-BERT, and SciBERT, with query expansion via pseudo-relevance feedback and WordNet synonyms, demonstrating that lexical methods can outperform neural alternatives on domain-specific biomedical corpora.

OpenMP Cholesky Factorisation Optimisation on HPC HPCC++OpenMP

Incremental optimisation of Cholesky matrix factorisation from a serial baseline to a tuned OpenMP parallel implementation, targeting the CSD3 HPC cluster. Progresses through loop reordering, reciprocal hoisting, persistent thread teams, and column scaling, with a comprehensive test suite validating numerical stability and reconstruction accuracy at each stage.

Calorimeter Energy Resolution Analysis Particle PhysicsStatistics

Statistical characterisation of particle detector response for a calorimeter: energy-dependent bias, resolution, and noise contributions across the full energy spectrum. Systematically compared six estimation methods — basic statistics with least-squares trending, per-energy unbinned MLE, simultaneous global likelihood fitting, profile likelihood confidence intervals, jackknife resampling, and bootstrap propagation — demonstrating that simultaneous fitting outperforms binned approaches by avoiding information loss.

High-Energy Physics: Gaussian Processes & Symmetry Invariants HEPGaussian Processes

Computational notebooks exploring two foundational techniques in high-energy physics analysis: Gaussian process regression for uncertainty-aware modelling of physical observables, and symmetry invariants — quantities preserved under physical transformations — applied to classification and regression tasks in particle physics contexts.

Galactic Archaeology — Stellar Stream Simulation AstrophysicsN-body

N-body simulation of tidal stellar streams using the Gala astrophysics library, with Palomar 5 as a case study. Employs the FardalStreamDF distribution function and a Plummer potential model, integrating particle orbits backward in time to reconstruct stream formation history and visualise particle distributions in galactic coordinates.

Pulse-Wave Velocity Analysis — Cardiovascular Signal Processing HealthTechSignal Processing

Research-grade pipelines for computing pulse wave velocity (PWV) from PPG (photoplethysmography) signals — a clinical marker of arterial stiffness and cardiovascular risk. Three independent pipeline implementations (SegmentalProcessing, TransitTimeMethod, CompositePWV) evaluated on both synthetic data and public PhysioNet datasets. Built with reproducibility as a first-class goal: GitHub Actions CI, pytest suite, CITATION.cff, and publication-quality figures.

FactTrace — Multi-Agent AI Fact-Checking System Winner: Best AI Product and Research 2026 Multi-agentLLMs

Multi-agent debate system where AI agents with distinct roles — skeptic, pedantic fact-checker, common-sense judge — argue and negotiate to reach a verdict on factual claims. Built in 24 hours at the Cambridge DiS Hackathon. Produces transparent reasoning chains rather than a binary true/false, surfacing ambiguity and uncertainty explicitly. Includes a visual HTML analysis dashboard and structured debate transcripts.

Interpoletor: 5D Regression System Full-StackFastAPINext.js

Production-grade platform for training, deploying, and querying PyTorch neural networks on five-dimensional regression tasks. Provides a REST API backend (FastAPI), a modern React/TypeScript frontend (Next.js 14), and full Docker containerisation. Features automated data validation, early stopping, real-time inference, and health checks.

LabSphere.ai — AI-Powered Lab Management System Winner: Best AI Project 2025

Award-winning multi-agent AI system for laboratory workflow automation, built at the Cambridge × OpenAI Hackathon 2025. Automates experiment scheduling, reagent inventory management, and resource allocation using multi-agent reinforcement learning, reducing manual coordination overhead for research teams.

Winner of Best AI Product and Research

University of Cambridge & FactTrace Inaugural Hackathon for creating a multi-agent system for AI truth verification

2026

Winner of Best AI Project

University of Cambridge & OpenAI Hackathon for developing fully functional AI-powered product

2025

Scholarship for International Studies

Association Solidarity FRANCE, French Government for Cambridge studies

2025-2026

Excellence Scholarship EUR PNGS-M&CS

Université Sorbonne Paris XIII (Issued by Graduate School in Mathematics and Computer Sciences)

2024 - 2025

Winner of The Best Innovative Research Project

International Competition of Scientific and Technical Research Projects: "Science in Motion: Discoveries in Engineering, Mathematics and Physics"

2023

Second Degree Diploma, The Best Innovative Research Project

International Competition: "Innovations in the Field of Information Technology"

2023

Winner of Universiade of the Faculty of Physics

Graduate Papers Competition in Lomonosov Moscow State University

2023

State Fellow for Special Academic Achievements

Lomonosov Moscow State University (Issued by Russian Government)

2020 - 2023