Theo Stathakopoulos

AI/ML Engineer with 3+ years of hands-on experience developing, deploying and monitoring end-to-end machine learning systems across LLM and agentic applications, computer vision, time-series forecasting and energy analytics. Experienced in building production-oriented AI solutions within GDPR and AI Act constraints, while translating technical results into clear business value for stakeholders. Strong background in evaluating AI solutions, synthetic data generation, real-time monitoring, MLOps, and cloud deployment (AWS SageMaker/Lambda, Google Cloud Run).

AI Engineer at Metlen (2025-04 – Present)

• Built a skills-driven AI agent (Google ADK + Gemini) that lets non-technical stakeholders ask questions in natural language about a 180-store energy dataset. It reasons step by step, loads skills on demand and writes its own SQL over a DuckDB analytics layer, inside a sandboxed read-only boundary with tool-callback guardrails and a verification step that rejects any figure not returned by an executed query. Built on a synthetic, GDPR-safe stand-in for the customer smart-meter data.
• Shipped the agent as an authenticated web service: FastAPI backend and Next.js chat UI with persistent, resumable multi-turn sessions; containerised and deployed on Google Cloud Run with Terraform (portable to Azure), behind Google IAP, with each user's chat history bound to their verified identity, not a client-supplied header.
• Designed and built the evaluation harness at two levels: each agent skill before it is promoted, and the whole solution end to end as a black box through the product's chat API, graded only on observable behaviour, with a guarded hold-out set. 11 automated checks - numeric answers verified against the database, tool trajectory and cost, safety, and an LLM-as-judge that only counts once calibrated - run over 211 test cases in 12 versioned datasets. An unmeasured case never counts as a pass, and every promotion goes through human review.
• Extended it to production: live Langfuse traces are scored on the same pipeline, a trace monitor screens them and flags suspicious conversations for LLM analysis, and token usage and cost are recorded per run.
• Developed the baseline energy-profiling pipeline for B2B smart-meter data (time-series forecasting and regression), reaching 5% daily MAPE under walk-forward validation; deployed on AWS SageMaker and Lambda with a local fallback for on-demand inference.
• Built a Streamlit and AWS QuickSight dashboard that tracks live model performance and drift, for earlier detection of degradation.

Machine Learning & Computer Vision Engineer (Researcher) at I-SENSE Group/ICCS, National Technical University of Athens (NTUA) (2024-04 – 2025-03)

• Built the back end of an IoT smart-meter system for real-time energy-consumption monitoring, from secure data acquisition to visualisation.
• Trained a 2D object-detection model and combined it with 2D-to-3D mapping over photogrammetric point clouds to estimate stockpile volumes and normalise terrain (EU project TERRAVISION).
• Expanded training data with Stable Diffusion pipelines (text-to-image, image-to-image, inpainting) for synthetic image generation and augmentation.
• Led MLOps: self-hosted MLflow and DVC, and automated model deployment with FastAPI and Docker through CI/CD.

Machine Learning & Computer Vision Engineer at Irida Labs (edge AI and computer vision) (2023-09 – 2024-02)

• Developed and optimised deep-learning models for real-time object detection, semantic segmentation and edge-AI deployment, tuning latency, FPS and model size.
• Built synthetic-data augmentation pipelines and foundation-model auto-labelling that sped up dataset preparation.
• Built Streamlit dashboards and FastAPI services for model interaction and evaluation.

Research Engineer at HEDNO - Hellenic Electricity Distribution Network Operator (DSO) (2022-09 – 2023-08)

• Led project activities and wrote deliverables for the EU H2020 project PlatOne: renewable-energy observability through PMU installation, and flexibility services in distribution grids.
• Developed and tested T-SQL scripts for energy-market clearing.

Mandatory military service at Dept. of Research & Computer Science (2021-09 – 2022-06)

• Maintained IT operations and built internal Python/SQL tools.

Machine Learning & Computer Vision Engineer (Internship) at Irida Labs (2021-03 – 2021-05)

• Applied GANs to day-to-night style transfer and evaluated the results with a YOLOv5 detector (PyTorch).

Master of Engineering in Electrical and Computer Engineering – University of Patras (2021-07)