Muberra Ozmen

Applied Scientist  ·  AI Safety Researcher

About

ML scientist with research and industry experience in graph learning and applied AI. I am pivoting to AI safety research, with recent work on LLM jailbreak detection, agent alignment evaluation, and graph-based risk detection. Currently an Applied Scientist at Coveo, where I work on multi-turn search agents and LLM evaluation pipelines.

Publications

  1. 8

    PsychoPass: Geometric Profiling of Multi-Turn Adversarial LLM Conversations

    M. Ozmen, S. Majumdar — arXiv preprint, 2026

  2. 7

    Understanding the Design Principles of Link Prediction in Directed Settings

    J. Zhai, M. Ozmen, T. Markovich — ACM WebConf Workshop (NEGEL), 2025

  3. 6

    Recent Link Classification on Temporal Graphs Using Graph Profiler

    M. Ozmen, T. Markovich — Transactions on Machine Learning Research (TMLR), 2024

  4. 5

    Benchmarking Edge Regression on Temporal Networks

    M. Ozmen, F. Regol, T. Markovich — Journal of Data-centric Machine Learning Research (DMLR), 2024

  5. 4

    Substituting Data Annotation with Balanced Neighbourhoods and Collective Loss in Multi-label Text Classification

    M. Ozmen, J. Cotnareanu, M. Coates — Conference on Lifelong Learning Agents (CoLLAs), 2023

  6. 3

    Multi-relation Message Passing for Multi-label Text Classification

    M. Ozmen, H. Zhang, P. Wang, M. Coates — IEEE ICASSP, 2022

  7. 2

    Microwave Radar for Breast Health Monitoring: System Performance Protocol

    L. Kranold, M. Ozmen, M. Coates, M. Popović — IEEE IMBioC, 2020

  8. 1

    Interactive Evolutionary Approaches to Multi-objective Feature Selection

    M. Ozmen, G. Karakaya, M. Koksalan — International Transactions in Operational Research (ITOR), 2017

Experience

Coveo Applied Scientist Montreal, QC, Canada
  • Designing a multi-turn LLM search agent integrated with Coveo's retrieval and indexing stack.
  • Built an evaluation and alignment pipeline to measure model adherence to behavioral guardrails and custom preference specifications.
  • Initiated the PsychoPass project: geometric profiling of adversarial multi-turn LLM conversations for early jailbreak detection.
Block Machine Learning Engineer Montreal, QC, Canada (Remote)
  • Led fine-tuning of the Foundational Events Model — a large-scale transformer encoding user activity sequences into user embeddings — for account takeover and chargeback detection.
  • Applied inductive GNNs at scale to identify gambling activity and referral abuse in transaction networks.
  • Developed anomaly detection on time series of customer location signals (Customer Journey).
Block Applied Deep Learning Intern Montreal, QC, Canada (Remote)
  • Formulated real-time chargeback detection as a temporal graph learning problem and developed a novel deep learning architecture for this task.
  • Work resulted in publications in TMLR and DMLR.
Segmentify Data Scientist Istanbul, Türkiye
  • Built a customer segmentation framework encoding purchasing behaviour for lifetime value estimation, supporting e-commerce retailers in boosting conversion rates.

Education

McGill University Ph.D., Computer Engineering Montreal, QC, Canada
University College London M.S., School of Management London, UK
  • Fully funded by UCL studentship.
Middle East Technical University M.S., Industrial Engineering Ankara, Turkey
  • Dean's Honour List; fully government-sponsored.
Middle East Technical University B.S., Industrial Engineering Ankara, Turkey
  • Dean's High Honour List.
Blue Dot Impact Certificate — Technical AI Safety Online
  • Intensive course on AI alignment, interpretability, and AI safety research methods. Verified certificate.

Skills

Languages
Python
Core ML
PyTorch PyTorch Geometric Transformers Lightning scikit-learn
Agentic / Safety
PyRIT MCP LangChain OpenAI API Anthropic API
MLOps
W&B MLflow Docker SLURM