Farid Ariai

Farid Ariai

PhD Candidate in Data Science | Legal NLP and LLM Evaluation

CV

Download my academic CV (PDF)

Profile

PhD candidate in Data Science at The University of Queensland, specialising in Legal NLP, LLM evaluation and efficient inference. Research spans legal retrieval-augmented generation, parameter-efficient fine-tuning, quantisation and human–AI interaction. Combines controlled experimentation with research software development, GPU/HPC workflows and human-in-the-loop evaluation.

Education

Research experience

PhD Researcher — Legal NLP and LLM Evaluation

The University of Queensland · April 2024–Present

  • Conduct research on legal retrieval, parameter-efficient model adaptation and post-training quantisation, with an emphasis on controlled evaluation and reproducible experiments.
  • Develop research software spanning dataset preparation, model training and inference, statistical evaluation and GPU/HPC execution.
  • Authored a survey of Legal NLP published in ACM Computing Surveys.

Research Collaborator — Digital Deliberative Democracy (D3)

The University of Queensland / University of Zurich · April 2024–Present

Contribute to research on NLP and human–AI interaction in civic participation through the D3 project, supported by SNSF Grant 205975.

Selected research projects

  • Co-developed a Legal RAG pipeline combining LLM-based information extraction, Neo4j, dense/BM25 retrieval, reranking and controlled tests of retrieval cues versus graph structure.

  • Contributed reproducible retrieval and evaluation infrastructure covering a 772,001-passage legal corpus and 4,584 HousingQA questions across 47 jurisdictions.

  • Co-developed controlled experiments on legal-entity supervision using QLoRA fine-tuning, vLLM inference, token-matched training and legal-citation counterfactuals.

  • Contributed to evaluation across CaseHOLD and MultiEURLEX, including two 7B model families, three seeds and a reconstructed 197-document legal-entity resource.

  • Built a configurable evaluation framework for six open LLM configurations from 7B to 72B parameters, comparing post-training quantisation from FP8 to sub-2-bit variants.

  • Executed and analysed 159 full-test result cells across LEDGAR, CaseHOLD and EURLEX, with paired statistical testing, structured-output audits, checkpoint-footprint measurement and backend controls.

LLM-Assisted Swiss Consultation Drafting

  • Developed an LLM-assisted consultation-drafting backend with FastAPI, PostgreSQL and SSE streaming, including participant-linked transcripts and usage telemetry.

  • Engineered transactional study-budget controls and recovery-aware multi-VM accounting, alongside representative multi-turn cost benchmarks and streaming load tests.

AI-Guided Online Deliberation

  • Built a reproducible analysis pipeline for a four-arm randomised online-deliberation experiment covering 336 participants, 8,118 comments and 411,515 cleaned behavioural events.

  • Implemented argument-tree compliance analysis, pre-registered statistical testing and LLM-based discourse annotation with blinded human validation and H100/Slurm inference.

LLM Code-Style Assessment and Human Evaluation

  • Developed a Python code-style assessment pipeline integrating assignment context, Pylint and schema-validated feedback across OpenAI, Anthropic and local vLLM workflows.

  • Built a shared 26-criterion human–AI annotation framework with source-integrity checks, reviewer adjudication and exact/overlap-aware evaluation of line-level findings.

Publications

Journal Articles

  1. Ariai, F., Mackenzie, J., & Demartini, G. (2025). Natural Language Processing for the Legal Domain: A Survey of Tasks, Datasets, Models, and Challenges. ACM Computing Surveys, 58(6), Art. 163, 1–37. doi:10.1145/3777009 Details
  2. Ariai, F., Tayefeh Mahmoudi, M., & Moeini, A. (2024). Enhancing Aspect-Based Sentiment Analysis with ParsBERT in Persian Language. Journal of AI & Data Mining, 12(1), 1–14. doi:10.22044/jadm.2023.13666.2482 Details
  3. Aryai, V., Kharazi, M., & Ariai, F. (2017). “Nearest Zero-point” Algorithm for Cooperative Robotic Search Missions. IAES International Journal of Robotics & Automation, 6(1), 49–58. doi:10.11591/ijra.v6i1.pp49-58 Details

Conference Papers

  1. Aryai, V., Mahmoodian, M., Vahidi Ferdowsi, N., & Ariai, F. (2018). Structural Reliability Analysis of Corroding Steel Bridges Using Random-Field Representation. Maintenance, Safety, Risk, Management and Life Cycle Performance of Bridges. 9th International Conference on Bridge Maintenance, Safety & Management (IABMAS 2018), Melbourne. Details

Teaching experience

Tutor — CSSE7030: Introduction to Programming
The University of Queensland, Brisbane · July 2025–Present

Tutor postgraduate students in Python programming, algorithmic thinking and software design.

Industry experience

Software Developer
Neyestan Rayaneh, Mashhad, Iran · September 2015–January 2024

  • Designed and maintained enterprise software systems.
  • Managed the full software-development lifecycle: requirements, implementation, testing and deployment.

Technical skills

  • Programming: Python, SQL, Bash, JavaScript, C++.

  • LLM adaptation and inference: PyTorch, Hugging Face Transformers, PEFT/LoRA, QLoRA, vLLM, llama.cpp.

  • Retrieval and evaluation: Dense retrieval, BM25, reranking, Neo4j/Cypher; FP8, AWQ and GGUF evaluation; structured-output and checkpoint-footprint analysis.

  • Applications and data systems: FastAPI, PostgreSQL, SQLite, SSE streaming, OpenAI and Anthropic API integration.

  • Experimental methods: Controlled comparisons, counterfactual evaluation, paired bootstrap intervals, multiple-testing correction, human annotation and agreement analysis.

  • Research infrastructure: Slurm, GPU/HPC workflows, tensor parallelism, Docker, resumable experiments and automated testing.

Languages: Persian (native); English (professional / academic proficiency).

Academic referees

Professor Gianluca Demartini — Principal Supervisor, School of EECS, UQ.
Dr Joel Mackenzie — Associate Supervisor, School of EECS, UQ.