Research

My research centres on Legal NLP and LLM evaluation: how language models use information, how efficiently they operate, and how their outputs and effects can be measured. Civic participation and educational AI provide complementary settings for this work.

Legal RAG: Retrieval Cues and Graph Structure

Legal retrieval research and experimental infrastructure

This project investigates whether the performance of graph-backed legal retrieval is explained by graph structure or by LLM-generated retrieval cues. It combines legal information extraction, Neo4j, dense and lexical retrieval, and controlled evaluation on bar-exam and housing-law questions, with particular attention to dataset provenance and the retriever’s actual behaviour.

MJL-NER: Legal Entity Supervision in LLMs

Controlled Legal NLP experiments and model adaptation

MJL-NER examines whether legal-entity supervision helps language models with legal question answering and document classification. The project combines QLoRA fine-tuning, vLLM inference and controlled evaluation on CaseHOLD and MultiEURLEX. Token-matched training conditions and citation-level counterfactuals test whether apparent effects reflect useful entity supervision, differences in training allocation, or associations that do not translate into model reliance.

Quantisation Limits of LLMs for Legal NLP

Efficient inference and empirical LLM evaluation

This project studies the quality–resource trade-offs of post-training quantisation for legal language tasks. It evaluates six core model configurations from 7B to 72B parameters across LEDGAR, CaseHOLD and EURLEX, from FP8 to sub-2-bit variants. The analysis goes beyond aggregate scores to examine structured-output validity, checkpoint size, task-specific failures and differences between inference backends.

LLM-Assisted Swiss Consultation Drafting

Experimental LLM platform and study infrastructure

This project develops research infrastructure for studying LLM-assisted participation in Swiss public consultations. The platform supports streamed, multi-turn conversations, participant-linked transcripts and detailed usage records. Its engineering includes transactional study-budget controls, recovery-aware multi-VM accounting, and representative cost and load testing, supporting a planned study of how AI assistance affects drafting and participant experience.

AI-Guided Online Deliberation

Experimental analytics, NLP measurement and computational social science

This project analyses how AI-selected argument nudges affect online political deliberation. Using a four-arm randomised experiment with 336 participants, the research pipeline combines surveys, 8,118 comments, behavioural logs and structured argument maps. It measures intervention uptake and discussion patterns, and integrates statistical analysis with LLM-based discourse annotation and blinded human validation.

LLM Code-Style Assessment and Human Evaluation

Educational AI evaluation and annotation tooling

This project investigates how LLM-generated feedback on Python code can be evaluated against human judgement. It combines assignment context, Pylint and schema-validated model outputs with a shared 26-criterion rubric. A human annotation application supports independent review and adjudication, while an evaluation engine compares model findings with consensus annotations using exact and overlap-aware line matching.