Marina Samprovalaki
CV

AI Engineer at Agile Actors Athens, Greece

I build agentic AI systems that hold up in production.

I design multi-agent workflows, retrieval pipelines and the evaluation that proves they work. That engineering is backed by published research in medical vision-language AI, including two 1st-place results at ImageCLEFmedical 2025.

Download CV
    running 0.0s
    Illustrative traces of three kinds of system I build: agentic RAG, diagnostic captioning and document structuring.

    01 Expertise

    Where I'm most useful

    Most AI demos work. Getting them to work reliably, on your data and in front of real users, is the hard part.

    01

    Agentic systems

    Multi-agent assistants and stateful workflows with planning, tool use and self-correction loops. I design them so failures are visible and recoverable, not silent.

    • LangChain
    • Multi-agent
    • Tool calling

    02

    Retrieval & agentic RAG

    Retrieval across documents, images and relational databases at once, with routing and reranking tuned to the question instead of a fixed top-k.

    • Multi-source RAG
    • Reranking
    • SQL + vectors

    03

    LLM evaluation

    Benchmarking accuracy and faithfulness, so a change ships because the numbers moved, not because a demo looked good.

    • Faithfulness
    • Benchmarks
    • Error analysis

    04

    Multimodal & vision-language

    Fine-tuning and steering vision-language models for medical imaging: captioning, concept-guided decoding and image–text reranking.

    • InstructBLIP
    • Guided decoding
    • MedCLIP

    05

    MCP & tool integration

    MCP servers exposing tools, resources and prompts over real data, reused unchanged across Claude, GPT, Gemini and open-weight models.

    • MCP
    • FastMCP
    • Function calling

    06

    Structured extraction

    Turning messy PDFs and Word files into typed, validated data using layout extraction, schema-constrained generation, deterministic checks and gated repair.

    • Pydantic
    • Azure Doc Intelligence
    • Jinja2

    03 Experience

    From full-stack to research to production AI

    1. 2024 — Now

      AI Engineer · Agile Actors

      • Build and ship a production multi-agent assistant that turns natural-language questions into SQL queries over a relational database.
      • Secure it with guardrails, including SQL validation, prompt hardening and attack monitoring.
      • Track SQL-generation accuracy with an LLM-as-judge evaluation pipeline I designed, and benchmark LLM systems for accuracy and faithfulness before they ship.
      • Architect multi-agent chatbots and stateful LLM workflows with LangChain and LangGraph.
      • Build agentic RAG pipelines for complex document and image retrieval across multiple databases.
      • Turn unstructured Word documents into structured Markdown with a pipeline built on Azure Document Intelligence.
      • Deploy and operate containerised services on OpenShift, including incident diagnosis and resolution.

    2. 2023 — 2025

      NLP Researcher · AUEB Natural Language Processing Group

      • Research diagnostic captioning with multimodal LLMs, using prompt engineering and retrieval augmentation to improve caption quality.
      • First author of the ImageCLEFmedical 2024 submission (2nd in Concept Detection).

    3. 2022 — 2023

      Junior Full Stack Engineer · HOMI Greece

      • Built classification and regression models for predictive tasks and deployed them on AWS SageMaker.
      • Shipped FastAPI services and React interfaces on top of PostgreSQL and MongoDB.

    2023 — 2024

    M.Sc. by Research, Computer Science

    Athens University of Economics and Business. Thesis supervised by Prof. Ion Androutsopoulos and Asst. Prof. John Pavlopoulos.

    2018 — 2022

    B.Sc. in Informatics

    Athens University of Economics and Business.

    04 Research

    Publications

    Working notes and thesis on biomedical vision-language modelling.

    1. CLEF 2025 Working Notes · Madrid

      AUEB NLP Group at ImageCLEFmedical Caption 2025

      A. Chatzipapadopoulou, I. Pantelidis, F. Charalampakos, M. Samprovalaki, G. Moschovis, P. Kaliosis, K. V. Dalakleidi, J. Pavlopoulos, I. Androutsopoulos

      Equal contribution among the first four authors · 1st in Concept Detection · 1st in Explainability

      PDF Code

    2. CLEF 2024 Working Notes · Grenoble · Highlighted talk

      AUEB NLP Group at ImageCLEFmedical Caption 2024

      M. Samprovalaki, A. Chatzipapadopoulou, G. Moschovis, F. Charalampakos, P. Kaliosis, J. Pavlopoulos, I. Androutsopoulos

      First author · Equal contribution among the first five authors · 2nd in Concept Detection · 4th in Caption Prediction

      PDF Code Slides Poster

    3. M.Sc. Thesis AUEB · November 2024

      Exploring Multimodal Large Language Models for Medical Image Captioning

      M. Samprovalaki. Supervised by I. Androutsopoulos and J. Pavlopoulos.

      Task-specific fine-tuning and few-shot learning of MLLMs for diagnostic captioning, and fusing alternative captions with a smaller language model.

      PDF Code Slides

    Currently exploring

    • Agentic reasoning & self-correction
    • Multi-agent orchestration
    • MCP & agent-to-agent protocols
    • Preference optimization (RLHF / DPO)
    • LLM evaluation
    • Foundation vision models

    05 Contact

    Building something with agents? Let's talk.

    I'm open to conversations about agentic systems, RAG, MCP and LLM evaluation, whether that's a role, a collaboration or a talk.

    marina.samprovalaki@gmail.com

    ↑↓ navigate↵ selectesc close