Yiorgos Papadopoulos | ML & LLM Systems Engineer

About

yiorgos@papadopoulos:~$whoami

name:: Yiorgos Papadopoulos
role:: ML & LLM Systems Engineer @ Zimmerman
studying:: MSc AI & Applications @ University of Thessaly
based in:: Thessaloniki, Greece
focus:: medical imaging, LLM systems, typed and tested Python backends
languages:: Greek, English

Projects

yiorgos@papadopoulos:~$ls projects/

thesis/: BSc thesis, graded 10/10
mammo/: Radiologist-in-the-loop decision support
model-pipeline/: ML service for grant-management analytics
narrative-engine/: Cited country narratives for the Global Fund
lead-gen/: Automated sourcing from IATI funding data

yiorgos@papadopoulos:~$cat projects/thesis.md

Chest X-ray pathology classifier

public, Write-up on GitHub

A deep learning system that reads chest X-rays, flags lung pathologies, and shows clinicians where it looked.

[ AUC 90%+ ]
[ 112K X-rays ]
[ Graded 10/10 ]

Seven disease-specific models, each fusing MobileNetV2 and ResNet50 backbones, trained on the NIH ChestX-ray14 dataset (about 112K images).
All models above 90% AUC, with two-phase training, mixed precision, and class weighting for imbalance.
Grad-CAM heatmaps for every prediction, so a clinician can check the model's reasoning.
A PyQt5 desktop app that runs inference and visualization end to end.

stack: Python · TensorFlow · Keras · OpenCV · Grad-CAM · PyQt5 · CUDA

            chest X-ray
                 |
         +-------+-------+
         v               v
  [ MobileNetV2 ]   [ ResNet50 ]
         |               |
         +-------+-------+
                 v
          [ fusion head ]
                 |
                 v
 per-disease score + Grad-CAM heatmap

yiorgos@papadopoulos:~$cat projects/mammo.md

Breast cancer screening assistant

Private repo, demo on request

A decision-support tool that flags, localizes, and prioritizes suspicious findings on screening mammograms, so radiologists read faster and miss less. The radiologist always makes the diagnosis.

[ Test AUC 0.844 ]
[ Patient-level splits ]
[ Cloud GPU training ]

VinDr-Mammo and CBIS-DDSM merged into one manifest, with patient-level splits and a leakage guard.
Separate models for masses (full image) and calcifications (high-resolution patches), with Grad-CAM localization.
Evaluated on AUC, sensitivity and specificity at an operating point, and calibration.
Clinical context checks: a breast-density masking flag and an image-adequacy check.
A FastAPI demo: upload a mammogram, get a score, a heatmap, and an urgent or not-urgent triage.

stack: Python · PyTorch · torchvision · FastAPI · pydicom · OpenCV · scikit-learn · Grad-CAM · uv

          DICOM / PNG
               |
               v
      [ adequacy check ]
               |
      +--------+-------+
      v                v
[ mass model ]   [ calc model ]
      |                |
      +--------+-------+
               v
   score + Grad-CAM heatmap
               |
               v
  triage: urgent / not urgent

yiorgos@papadopoulos:~$cat projects/model-pipeline.md

Model Pipeline

private, Zimmerman case study

Made 24 machine learning use cases in a production analytics service correct, reproducible, and small.

[ 24 use cases ]
[ GBs to ~1MB ]
[ 1,300+ tests ]
[ 100% coverage ]

Audited models across regression, classification, clustering, anomaly detection, optimization, network analysis, and causal inference.
Removed target leakage, fixed aggregation and train/test splits, and aligned prediction inputs with production data.
Capped model size, cutting saved artifacts from several GB to about 1MB.
Added experiment tracking (config, winning model, hyperparameters, data and git versions) and deterministic run replay.
Drove the service to 100% test coverage behind ruff, flake8, bandit, and type checking.

stack: Python · Flask · scikit-learn · Pandas · Pydantic · pytest · Swagger

 data -> features -> train and select
                          |
                          v
                     tracked run
                          |
                          v
                      REST API

yiorgos@papadopoulos:~$cat projects/narrative-engine.md

Narrative Engine

private, Zimmerman case study

Turns Global Fund country data and supporting documents into validated, cited written narratives with LLMs.

[ Cited outputs ]
[ Cost-guarded ]
[ 100% coverage ]

Collects country evidence and supporting documents, then generates narratives through LiteLLM with validation and citations.
Generation is an explicit, cost-guarded job with a persistent spend ledger, so model costs never surprise anyone.
Results are saved as versioned SQLite bundles; the API only serves saved content, so visitors never trigger a model call.
Feature-flagged integration with the Data Explorer frontend, behind 100% test coverage.

stack: Python · Flask · LiteLLM · Pydantic · SQLite · Docker · pytest

 evidence + documents
          |
          v
       LiteLLM
          |
          v
      validator
          |
          v
  versioned bundle (SQLite)
          |
          v
    read-only API
          |
          v
    Data Explorer

yiorgos@papadopoulos:~$cat projects/lead-gen.md

AI Lead Generation Agent

private, Zimmerman case study

Replaced a manual sales-research process with a pipeline that finds and qualifies leads in international aid funding data.

[ ~875K activities ]
[ 5 detectors ]
[ Cost-tracked LLM calls ]
[ 100% coverage ]

Queries the IATI Datastore (Solr, about 875K activities) to source and qualify leads.
Five organisation-level signal detectors: new funding, first-time publisher, activity spike, dormant publisher, and large commitment, each returning structured evidence.
Three-stage contact enrichment (cached runs, then IATI data, then OpenAI web search as a fallback) that cuts paid LLM calls, with per-run token and cost tracking.
A month-by-month historical replay to measure lead volume, at 100% test coverage.

stack: Python · Flask · Solr · OpenAI API · Pydantic · pytest

 IATI Solr (~875K activities)
          |
          v
 [ 5 signal detectors ]
          |
          v
       filters
          |
          v
 enrichment: cache > IATI > LLM
          |
          v
   qualified leads

Experience

yiorgos@papadopoulos:~$git log --author="Yiorgos"

commit aaab150 (HEAD -> main)

Date: Feb 2026 to present

ML & LLM Systems Engineer @ Zimmerman

Building backend services that integrate ML, NLP, LLMs, and embedding-based retrieval, exposed through typed APIs and React frontends.
Owning backend features end to end, from design through PR review to deployment.
Working closely with senior engineers through multi-round code reviews on GitHub.
Shipping in agile sprints and presenting completed work to the engineering team.

Stack

yiorgos@papadopoulos:~$cat stack.toml

[machine_learning]

items = [

"PyTorch", "TensorFlow", "Keras", "scikit-learn", "Pandas", "NumPy", "OpenCV", "Grad-CAM",

]

[llms_and_retrieval]

items = [

"LiteLLM", "OpenAI API", "embeddings", "semantic search", "prompt evaluation",

]

[backend]

items = [

"Python", "Flask", "FastAPI", "Pydantic", "REST and Swagger", "SQL", "MySQL", "SQLite", "Solr",

]

[engineering]

items = [

"Git and GitHub", "code review", "pytest", "100% coverage", "pre-commit", "ruff", "Docker", "uv", "Jira",

]

[languages]

items = [

"Python", "SQL", "C++", "JavaScript", "HTML",

]

Education

yiorgos@papadopoulos:~$cat education.md

MSc in Artificial Intelligence and Applications

University of Thessaly · Oct 2026 to present

BSc in Computer Science and Telecommunications

University of Thessaly · 2021 to 2026

GPA 7.5/10
Specialization in Data Management and Artificial Intelligence
Thesis graded 10/10

Erasmus+ exchange

ISLA Polytechnic Institute, Porto · 2024 to 2025

certifications:

AI in Healthcare Specialization, Stanford University
Generative AI for Healthcare, Google Cloud (2026)

Try it

Welcome. You can type here.

Try help, or pick a command from the list.