Patrik P. Süli

Software architect by day, applied‑LLM researcher by night.
PhD candidate at Obuda University; founder of KenguTech.

§ profile

A decade of professional software engineering, primarily on the JVM: Spring Boot, Quarkus, modern Java, and microservice systems sized for millions of requests. Today that experience funds and informs the second career, a PhD on applied LLMs at Obuda University's Doctoral School of Applied Informatics and Applied Mathematics.

My research sits at the intersection of large language models and applied science: agentic systems for research workflows, LLM-as-judge auditing, and neuro-symbolic approaches. Selected publications and working papers are listed under Research.

I also teach at Obuda University's John von Neumann Faculty of Informatics and run KenguTech, a small studio for production engagements where seniority actually matters. The two careers do not compete; they sharpen each other.

§ industry experienceFull timeline
  1. Sep 2026– Present
    HUN-REN Centre for Agricultural Research (ATK)Software Architect · AI Engineer
  2. Dec 2022– Present
    KenguTechSoftware Architect · Founder
  3. Aug 2024– Present
    Obuda University · John von Neumann Faculty of InformaticsAssistant Lecturer & Researcher
  4. Jan 2018– Nov 2022
    FreelanceSoftware Developer · Consultant
  5. Jul 2016– Jun 2019
    NETI IT Consulting Ltd.Software Developer · Trainee Developer
§ public codeAll projects
  1. Co-author
    benchgaphollorol/benchgapVisit benchgap.net
  2. Author
    DistribirdHUN-REN-AI1Science/Distribird
  3. Contributor
    SpotBugsspotbugs/spotbugs
§ find me
  1. Email
  2. Phone
  3. LinkedIn
    patrik-süli
§ currently writing01 / 06
in progress2026· In preparation

benchgap

Mind the Gap: Explicit Cross-Benchmark Calibrations for Estimating Missing LLM Evaluation Scores

Most LLMs are evaluated on only a few benchmarks: across 401 models and 146 benchmark versions, coverage is just 8.3%. benchgap estimates the missing scores without running any evaluations, calibrating each benchmark against same-capability benchmarks with monotone curves selected by leave-one-out cross-validation, and attaching an error and confidence level to every estimate. In a 15% holdout it fills 71% of masked cells at 4.6 pp MAE, 23% lower than a matrix-completion baseline. It also quantifies the harness tax, the disagreement between evaluation setups on the same models. Estimates are updated daily at benchgap.net.

Süli, P. P., Eigner, G., Hollós, R.

With Obuda University · Biomatics & Applied AI Institute · HUN-REN Centre for Agricultural Research

Try the toolSource
§ selected works9 INDEXED · 6 WORKING
See all research
  • 2026INESfirst author

    Süli, P. P., Eigner, G., Zrubka, Z.

    Emergent Health Preferences in Large Language Models: A Time-Trade-Off Analysis Using EQ-5D-3L

    Runs the standard time-trade-off protocol on five LLMs across all 243 EQ-5D-3L health states under 31 demographic personas (165,726 responses), comparing model-derived utilities to population norms and characterising model-specific biases.

    Type
    Conference paper
    Venue
    IEEE 30th Jubilee International Conference on Intelligent Engineering Systems (INES 2026), Budapest, Hungary
    DOI
    10.1109/INES69513.2026.11661097
  • 2025CINTIfirst author

    Süli, P. P., Tafferner-Gulyás, V., Kovács, L., Eigner, G.

    A Survey of Large Language Models as Research Assistants

    Survey of LLM-as-research-assistant patterns: literature triage, drafting, evaluation, and the failure modes that still demand human oversight.

    Type
    Conference paper
    Venue
    IEEE 25th International Symposium on Computational Intelligence and Informatics (CINTI 2025), Budapest, Hungary
    DOI
    10.1109/CINTI67731.2025.11311841
  • 2025ICGALfirst author

    Süli, P. P., Marczisák, B., Csibrák, T., Eigner, G., Zrubka, Z.

    Augmenting virtual brainstorming workshops via agentic AI

    Agentic-AI augmentation for virtual brainstorming workshops: coordinating prompt agents, idea consolidation, and human-in-the-loop ranking.

    Type
    Abstract
    Venue
    International Conference and Summer Institute on GenAI Literacy (ICGAL 2025), Budapest, Hungary