LLMs · AI for Science · Document AI

Islam Mesabah إسلام مصبح

PhD student @ RPTU Kaiserslautern-Landau · Researcher @ DFKI

Teaching AI to do research.

I build LLM systems that turn messy documents into structured data, and study how AI agents reason, collaborate and, eventually, do research on their own. Every year I teach generative-AI engineering at RPTU.

AI for research. An AI agent moves through the research loop: read, hypothesise, experiment, analyse, write.

Agentic AI. Specialised LLM agents plan, act, use tools and review each other.

Document AI. OCR and LLMs turn a scanned receipt into a structured record.

Code generation & optimisation. Documentation-grounded code generation, then measured speed-ups with every test passing.

Vision-language reasoning. A vision-language model classifies a traffic sign by walking an explicit decision tree.

NLP. A code-switched word is split into its Arabic and English parts, and each part is tagged with its language.

papers & preprints
6
citations · h-index 3 (Scholar, Sept 2026)
51
theses supervised
7
open-source releases & contributions: code, datasets, models
7

About

Who I am

Portrait of Islam Mesabah
Kaiserslautern, Germany
North Sinai Cairo Kaiserslautern
  • PhD student in Computer Science at RPTU Kaiserslautern-Landau (since Dec 2025), working on LLMs for code generation and optimisation, towards AI that can do scientific research, supervised by Prof. Sebastian Vollmer and Dr. Brian Moser.
  • Researcher at the German Research Center for Artificial Intelligence (DFKI), Department of Data Science and its Applications led by Prof. Sebastian Vollmer, since May 2023.
  • I work on large language models for document understanding, code generation and multi-agent collaboration, and on trustworthy, responsible AI.
  • Lecturer and coordinator of the Engineering with Generative AI course at RPTU.
  • M.Sc. Computer Science, TU Kaiserslautern (2024). B.Sc. Computer Science and Engineering, German University in Cairo (2021), where I also taught for a year.
  • Languages: Arabic (native), English (advanced), German (beginner)

Interests Large language models · Natural language processing · AI for scientific research · Document understanding and key-information extraction · Code generation and optimisation · Agentic and multi-agent AI · Vision-language reasoning · Trustworthy and privacy-preserving AI

Research

What I work on

AI for Science

Teaching AI to do research

Can AI do research, not just assist with it? My PhD builds LLM systems that write, run, evaluate and optimise code in a loop, then grows them into automated research workflows where experiment design, implementation and evaluation refine each other. The crux is trust: the system has to weigh scientific evidence, not just produce plausible output. My first step, ConflictBench, tests whether AI can detect, classify and explain disagreements between studies across seven research domains.

Document AI

Reading messy documents with OCR and LLMs

Turning scanned receipts, forms and PDFs into structured records. GAIDA combines OCR with LLM pre-annotation and a human in the loop, cutting annotation time by up to 80% and producing the AraReceipt dataset of Arabic receipts. We are now applying this to German building data, turning messy building-asset records into a structured database register.

NLP

Arabic–English language and text style

Natural language processing for mixed Arabic–English text: identifying the language of each part of a code-switched word, and evaluating how well large language models transfer text style.

Code generation & optimisation

Grounding code generation in API documentation

CoderAPI fine-tunes Instruct CodeT5+ 16B and retrieves API documentation at generation time so the model writes correct multi-line Python for both seen and unseen libraries. I also explore LLMs that optimise existing code, and contributed a prompt-optimisation example to the open-source evolutionary coding agent OpenEvolve.

Agentic AI and reasoning

How LLM agents collaborate, compete and reason

Networked multi-agent games where a story primes how agents cooperate, and tree-based reasoning that decomposes what vision-language models actually do when they classify an image.

Trustworthy AI

Keeping private data out of the prompt

Privacy Guardrail is a local-first browser extension that detects and anonymises personally identifiable information before text is pasted into ChatGPT, Claude or Gemini. Recognition runs entirely in the browser, with Rust/WebAssembly matchers and an optional transformer NER model, so nothing leaves the machine to be checked.

Experience

Experience and education

Experience

  1. May 2023 – Present

    Researcher, teaching assistant and PhD student

    German Research Center for Artificial Intelligence (DFKI) · Kaiserslautern, Germany

    • Structuring documents with OCR, LLM prompting and deep learning.
    • LLM tuning and validation; agent collaboration; LLMs for classification, code and research.
    • Responsible and trustworthy AI projects.
    • Coordinate and teach the annual Engineering with Generative AI course.
  2. Aug 2022 – May 2023

    Research assistant

    TU Kaiserslautern (now RPTU), Machine Learning Group · Kaiserslautern, Germany

    • Evaluating large language models (GPT-3, OPT, BLOOM) on English text style transfer.
  3. Sept 2021 – Jun 2022

    Research and teaching assistant

    German University in Cairo · New Cairo, Egypt

    • Taught Database Systems and Concepts of Programming Languages; supervised four bachelor theses.
  4. Aug 2021 – Oct 2021

    Backend developer intern

    Robusta Studio · Cairo, Egypt

    • Built a real-time Laravel API with validation, authentication and MySQL persistence in an agile team.

Education

  1. Dec 2025 – Present

    Ph.D. in Computer Science

    RPTU Kaiserslautern-Landau · Kaiserslautern, Germany

  2. Apr 2022 – Mar 2024

    M.Sc. in Computer Science

    TU Kaiserslautern (now RPTU) · Kaiserslautern, Germany

    • Specialisations in Intelligent Systems and Information Systems.
    • Thesis: Large language models for instructed and effective code generation using API documentation (CoderAPI).
    • Supervisors: Prof. Sebastian Vollmer and Dr. Sergey Redyuk.
  3. Oct 2016 – Jul 2021

    B.Sc. in Computer Science and Engineering

    German University in Cairo · New Cairo, Egypt

Publications

Papers

6 of 6 publications · 2021–2026

Show
Role
Type
Year

2026

  1. Venue TBA to appear 2026 Selected publication

    GAIDA: Generative AI for Human-in-the-Loop Document Annotation

    Islam Mesabah, Ali Elhosary, Priyabanta Sandulu, Darko Obradovic, Sebastian J. Vollmer

    2026 (venue to be announced)

    OCR plus LLM pre-annotation with a human reviewer cuts document-annotation time by up to 80% and yields AraReceipt, a dataset of 100 Arabic receipts with 25 key-information classes.

    BibTeX
    @inproceedings{mesabah2026gaida,
      title     = {{GAIDA}: Generative {AI} for Human-in-the-Loop Document Annotation},
      author    = {Mesabah, Islam and Elhosary, Ali and Sandulu, Priyabanta and Obradovic, Darko and Vollmer, Sebastian J.},
      year      = {2026}
    }

2025

  1. arXiv 2025 Selected publication

    Decomposing Visual Classification: Assessing Tree-Based Reasoning in VLMs

    Sary Elmansoury, Islam Mesabah, Gerrit Großmann, Peter Neigel, Raj Bhalwankar, Daniel Kondermann, Sebastian J. Vollmer

    arXiv preprint arXiv:2509.09732

    Does forcing a vision-language model to classify through an explicit decision tree help or hurt? We measure where tree-based reasoning adds accuracy and interpretability, and where it breaks.

    BibTeX
    @article{elmansoury2025decomposing,
      title   = {Decomposing Visual Classification: Assessing Tree-Based Reasoning in {VLMs}},
      author  = {Elmansoury, Sary and Mesabah, Islam and Gro{\ss}mann, Gerrit and Neigel, Peter and Bhalwankar, Raj and Kondermann, Daniel and Vollmer, Sebastian J.},
      journal = {arXiv preprint arXiv:2509.09732},
      year    = {2025}
    }
  2. NETGCOOP 2025 Selected publication

    The Power of Stories: Narrative Priming in Networked Multi-Agent LLM Interactions

    Gerrit Großmann, Larisa Ivanova, Sai Leela Poduru, Mohaddeseh Tabrizian, Islam Mesabah, David A. Selby, Sebastian J. Vollmer

    NETGCOOP 2025, Springer LNCS (also presented at the 1st International Electronic Conference on Games)

    Priming LLM agents with different stories changes how they cooperate in networked public-goods games: shared narratives raise contributions, mixed narratives fracture them.

    BibTeX
    @inproceedings{grossmann2025stories,
      title     = {The Power of Stories: Narrative Priming in Networked Multi-Agent {LLM} Interactions},
      author    = {Gro{\ss}mann, Gerrit and Ivanova, Larisa and Poduru, Sai Leela and Tabrizian, Mohaddeseh and Mesabah, Islam and Selby, David A. and Vollmer, Sebastian J.},
      booktitle = {Networks, Games, Control and Optimization (NETGCOOP 2025)},
      series    = {Lecture Notes in Computer Science},
      publisher = {Springer},
      year      = {2025},
      doi       = {10.1007/978-3-032-09315-8_11}
    }

2023

  1. ICICIS 2023

    An Intrusion Detection System for Smart Autonomous E-Bikes

    Roba Maged, Mohamed Sabry, Ahmed Ali, Islam Mesabah, Ahmed Mazhr, Hassan Soubra

    11th International Conference on Intelligent Computing and Information Systems (ICICIS), IEEE, pp. 233–240

    A lightweight intrusion-detection layer for connected e-bikes that flags tampered sensor and control messages before they reach the motor controller.

    BibTeX
    @inproceedings{maged2023intrusion,
      title     = {An Intrusion Detection System for Smart Autonomous E-Bikes},
      author    = {Maged, Roba and Sabry, Mohamed and Ali, Ahmed and Mesabah, Islam and Mazhr, Ahmed and Soubra, Hassan},
      booktitle = {2023 Eleventh International Conference on Intelligent Computing and Information Systems (ICICIS)},
      pages     = {233--240},
      publisher = {IEEE},
      year      = {2023},
      doi       = {10.1109/ICICIS58388.2023.10391189}
    }
  2. CPERE 2023

    Virtual Bike Lanes for Smart, Safe, and Green Navigation

    Mariam Sherif Seoudi, Islam Mesabah, Mohamed Sabry, Hassan Soubra

    IEEE Conference on Power Electronics and Renewable Energy (CPERE), pp. 1–6

    Route cyclists through virtual lanes chosen for safety and emissions rather than distance, using live traffic and air-quality data.

    BibTeX
    @inproceedings{seoudi2023virtual,
      title     = {Virtual Bike Lanes for Smart, Safe, and Green Navigation},
      author    = {Seoudi, Mariam Sherif and Mesabah, Islam and Sabry, Mohamed and Soubra, Hassan},
      booktitle = {2023 IEEE Conference on Power Electronics and Renewable Energy (CPERE)},
      pages     = {1--6},
      publisher = {IEEE},
      year      = {2023},
      doi       = {10.1109/CPERE56564.2023.10119537}
    }

2021

  1. Array 2021 Selected publication

    Language Identification of Intra-Word Code-Switching for Arabic–English

    Caroline Sabty, Islam Mesabah, Özlem Çetinoğlu, Slim Abdennadher

    Array, vol. 12, 100104, Elsevier

    Words like "elgame" mix Arabic and English inside a single token. We segment such words and label each part's language with a subword-level neural model.

    BibTeX
    @article{sabty2021language,
      title   = {Language Identification of Intra-Word Code-Switching for {Arabic}--{English}},
      author  = {Sabty, Caroline and Mesabah, Islam and {\c{C}}etino{\u{g}}lu, {\"O}zlem and Abdennadher, Slim},
      journal = {Array},
      volume  = {12},
      pages   = {100104},
      year    = {2021},
      publisher = {Elsevier},
      doi     = {10.1016/j.array.2021.100104}
    }

Academic service

Reviewing and committees

Awards

Scholarships and prizes

  1. Deutschlandstipendium

    German Federal Government and AWS

    National scholarship for high-performing students, awarded for academic achievement.

  2. Al-Thuraya MENA Scholarship

    Al-Thuraya Foundation

    Partial scholarship for the M.Sc. in Computer Science at TU Kaiserslautern.

  3. Second place, HackTrick Hackathon 2022 (AI track)

    Dell Technologies

    Led a team of four to second place and a 2,000 USD prize.

  4. First place, Technical Innovation Award, Shell Eco-Marathon Asia 2019

    Shell, Malaysia

    With GUC Innovators, a drive-by-wire system controlled through brainwave and facial sensors for drivers with physical limitations.

  5. Full scholarship for the B.Sc. in Engineering

    German University in Cairo

    Awarded for ranking among the top secondary-school students in the North Sinai governorate.

Community

Volunteering

  • Feb 2023 – Present

    Egypt Scholars Inc.

    DAR team member · California, USA (remote)

    Organise and host online educational webinars, from inviting the speaker to publishing the recording. Seven webinars published so far.

  • Oct 2019 – Jul 2020

    IEEE GUC Student Branch

    Software committee member · New Cairo, Egypt

    Ran sessions explaining software topics to committee members.

  • Dec 2018 – Aug 2021

    GUC Innovators

    Research and innovation team head (2020–21), member (2018–20) · New Cairo, Egypt

    Student engineering team competing at the Shell Eco-Marathon Asia. Led the R&I team and built the team's Android app backed by Firebase.

    Team videos

Skills

How my skills connect

49 skills · 129 connections

Every node is a tool or language I work with. Two nodes are joined when they show up in the same paper, research thread, course or job. Select a node to see where I used it.

Python TypeScript JavaScript Rust Java C / C++ SQL Julia PHP PyTorch Hugging Face Transformers DeepSpeed TensorFlow Keras scikit-learn NVIDIA CUDA NumPy pandas LangChain AutoGen Dify Neo4j Amazon Bedrock Gemini CLI Cursor PaddleOCR OpenCV Label Studio CVAT Docker GPU clusters (Slurm) AWS Lambda · API Gateway · DynamoDB MySQL MongoDB Firebase React / MERN Node.js WebAssembly Laravel Astro HTML5 CSS Git GitHub GitLab Jupyter Google Colab Unity LaTeX
Select a node to see where it was used. Press Escape to clear.