01Experience

  1. Apr 2025
    — Now

    Software engineer at

    GenAI Engineering team. Backend tooling and infrastructure for LLM-backed product features. Production systems, not research prototypes.

  2. 2023 — 2025

    MSc Computer Science, and UvA

    Joint programme. Coursework in distributed systems, cloud computing, software architecture. Thesis on MLOps pipelines for CNC anomaly detection (MLflow vs. Kubeflow), with .

  3. 2021 — 2025

    Software engineer, PhoenixNAP

    Java engineer at a bare-metal cloud provider. Automated RAID configuration, custom OS image creation, Spring Boot provisioning tools, Temporal workflow migration, iPXE / SSH / Ansible integration, MongoDB-based distributed locking.

  4. 2018 — 2021

    Software engineer intern, CCBill

    A PhoenixNAP brand. Three summers building a full-stack internal employee management tool with a small team (Java, Spring, Maven, SQL, JS), through to deployment and support.

  5. 2018 — 2021

    BSc Artificial Intelligence, University of Malta

    Machine learning, computer vision, knowledge representation. Dissertation on saliency-directed product placement.

02Selected work

  1. Key finding·CAIN 2026
    No single tool covers the full MLOps lifecycle. Teams stitch several together — interoperability becomes the central concern.
    41papers reviewed across orchestration, versioning, tracking, deployment, monitoring.
    Peer-reviewed paper

    A Systematic Review of MLOps Tools

    Micallef, Z., Rajenthiram, K., and Gerostathopoulos, I. (2026). In Proceedings of the 5th International Conference on AI Engineering (CAIN '26). ACM, Rio de Janeiro.

    Maps MLOps-native tools to lifecycle components, identifies adoption patterns, and synthesises reported limitations from real-world usage. Fed directly into the MSc thesis pipelines built on 's industrial use case.

    DOI →ResearchGate →
  2. Industrial MLOps pipeline figure
    MSc Thesis

    Industrial MLOps for Anomaly Detection

    Two pipelines (one MLflow, one Kubeflow) built around an industrial anomaly-detection use case on CNC machine signals, with . Honest finding: MLflow gets you running quickly; Kubeflow needs Kubernetes fluency before it gives you anything. Both work — the choice is mostly about what the team already runs.

    Thesis PDF →Lit. Review →Zenodo →
  3. Saliency-directed product placement
    BSc Dissertation

    Saliency-Directed Product Placement

    Computer-vision system that ranks products in a scene by predicted attention. Mask R-CNN object detection plus a custom saliency-segment ranking algorithm. 0.66 correlation with human attention on a held-out set. Python, OpenCV, Mask R-CNN, Supervisely.

    GitHub →Report →Demo →