AI Automation
Repetitive work — document handling, reporting, data entry, triage — rebuilt so software does it instead of a person. I look at the boring 80% first, because that is where the payback is.
Hello, I'm
George Levis
AI/ML Engineer
I build working systems, not slide decks. My day job is an enterprise agentic AI platform — LangGraph, MCP servers, autonomous LLM agents running on real data infrastructure. Alongside it I help companies put AI, automation and plain good software on actual problems: processes that eat hours every week, tools that don't talk to each other, work that shouldn't need a person. If it ships and holds up in production, I'm interested.
Interests
Deep LearningReinforcement LearningLLMsFederated LearningMLOpsRAGMulti-Agent Systems
Technology Stack
Python
PyTorch
Docker
AWS
K8s
Git
Linux
SQL
Services
What I Do
I help companies solve concrete problems with AI, automation and software — by building the thing, not just recommending it.
AI Integration
Connecting AI to the systems you already run — databases, internal tools, CRMs — through proper interfaces rather than another disconnected app. This is my day job: MCP servers and agent tooling over enterprise data infrastructure.
AI Training
Hands-on sessions for teams that want to use Claude, ChatGPT and similar tools properly: what to delegate, what to verify, and how to prompt so the output holds up. Based on how I use these tools daily, not on vendor slides.
AI Agents & Workflows
Agents that plan, call tools and finish multi-step work on their own — the kind I build at Cognity with LangGraph and agent SDKs. For when a fixed script is not enough and a person in the loop is too slow.
Websites & Web Apps
Fast, maintainable sites and web applications — Vue/Nuxt, TypeScript, deployment that runs itself. This site is one of them: statically generated, CMS-backed, rebuilt on every push.
Custom Software
The internal tool your team needs but cannot buy: dashboards, APIs, document Q&A services, the glue between two systems that refuse to talk. Built small, documented, and handed over so you are not locked in.
Machine Learning & Data
When the problem genuinely needs a model — forecasting, anomaly detection, a fine-tuned LLM — I have built and shipped them, with peer-reviewed research behind the theory. I will also tell you when it does not need one.
Something else?
Most real problems do not fit a tidy box. Describe yours and I will tell you what I would build — or that you do not need me.
Get in touch →Portfolio
Things I've Built
DocuRAG
A Python-based Retrieval-Augmented Generation (RAG) system for document question-answering. Built with FastAPI, HuggingFace transformers, and FAISS vector search. Supports PDF, TXT, DOCX, MD ingestion with both REST API and Streamlit web interface.
PythonRAGFastAPI
Smart Home P2P Energy Trading with RL
A deep reinforcement learning approach to optimize smart home energy usage while preventing cartel-like behavior in peer-to-peer energy markets. Implements DDPG to optimize HVAC operation, battery management, and price-setting strategies with anti-cartel mechanisms for market fairness.
PythonPyTorchDDPG
Finetuning LLM Gemma3 for AI Detection
A state-of-the-art AI text detection system built by fine-tuning Google's Gemma3-4B model using Unsloth optimizations and the RAID dataset. Detects AI-generated text across 11 LLMs and multiple domains with memory-efficient training on consumer hardware.
PythonLLMFine-tuning
Research
Featured Publication
Multi-agent deep reinforcement learning for peer-to-peer energy trading in smart home communities — balancing comfort, cost, and market fairness.
All PublicationsFedRA: A Fully Decentralized Federated Learning Framework for Robust Intelligence Sharing across O-RAN dApps
Sotirios T. Spantideas, Anastasios E. Giannopoulos, George A. Levis, Panagiotis Trakadas
We propose a resilient and fully decentralized federated learning framework specifically adapted for Open Radio Access Networks (O-RAN) named FedRA that enables collaborative intelligence between dApps. We present how FedRA can be deployed in the O-RAN architecture by using already existing interfaces between the well-defined RAN components and how real-time ML applications (dApps) can interact to share their distilled intelligence. The software components of FedRA are described along with the step-by-step deployment workflow, outlining practical guidelines for its implementation. Finally, the framework is validated in a simulated environment using both real and simulated datasets reporting the time series of throughput provided by multiple radio units; two different ML models hosted in FedRA nodes either detect anomalies in upcoming network traffic or forecast data-rate values. The achieved accuracy confirms the validity of FedRA in the training of cell-specific ML models and the decentralized collaborative intelligence sharing among them. FedRA is also compared to a typical client/server approach in terms of model accuracy and network communication cost.
Have a real problem to solve?
A process that eats hours every week, systems that refuse to talk to each other, an idea that needs to become software — tell me about it. I'll give you a straight answer on whether AI actually helps, what I'd build, and what isn't worth the effort.
Get in Touch