Kostas Psychogios - Data Science Contractor

Recommended expert

Athens, Greece

Experience

Apr 2025 - Dec 2025

9 months

Amsterdam, Netherlands

Created an evaluation pipeline for a RAG system using an LLM-as-a-judge approach, incorporating answer quality metrics such as comprehensiveness and harmfulness, as well as custom reproducibility and consistency metrics.
Implemented workflow enhancements in the existing RAG system, expanding business case coverage and increasing customer sales by 20%.
Improved the existing RAG model by adding error-handling branches to ensure a valid answer is always returned.
Collaborated with subject matter experts to design evaluation datasets for key components of the pipeline, including question rephrasing, document retrieval, and answer generation.

Jan 2024 - Apr 2025

1 year 4 months

Athens, Greece

Developed a RAG-based agent system for complex files (PDF, Word, scanned images, often multilingual) that accurately retrieved and generated relevant information for users using OCR, segmentation, and vision LLM models to handle unstructured documents containing figures, tables, and scanned pages.
Fine tuned an embedding model with a custom dataset and labels, achieving higher accuracy compared to the baseline with Hugging Face on a multi-GPU setup.
Operated in an agile environment with bi-weekly sprints and regular daily/weekly meetings.
Utilized Docker to deploy models and APIs, and employed Kubernetes to manage Google Cloud resources.

May 2022 - Jan 2024

1 year 9 months

Athens, Greece

Implemented an intrusion detection system trained within a federated learning framework, achieving an F1-score of 95% in malicious packet detection using TensorFlow and TensorFlow Federated.
Developed a local style transfer technique using Facebook’s Segment Anything Model, achieving visually captivating results.
Employed software engineering tools and frameworks such as Docker, FastAPI, Flask, and Git to ensure efficient project reuse and version control.

Aug 2021 - May 2022

10 months

Athens, Greece

Collaborated with two engineers to implement a sales forecasting system that accurately predicted weekly sales quantities for over 100 products using statsmodels, scikit-learn, and Keras.
Developed an NLP-based sentiment analysis module utilizing TensorFlow and NLTK, achieving a 90% accuracy rate in classifying customer comments in real-world scenarios.
Built and deployed a churn prediction model that reduced client operational costs by 20% through accurate user classification.