Eduard-David
Gyarmati

AI Solutions Engineer

01

Work

Kibo AI

AI Solutions Engineer

Nov 2025 – Present Nürnberg, DE
  • I build LLM-powered production systems end-to-end: API integration, prompt engineering, error handling, deployment
  • I prototype LLM integrations for clients, moving from rough proof-of-concept to deployed internal tools
  • I benchmark new models and frameworks against real use cases, not just leaderboard scores
Python TypeScript PyTorch Anthropic API PostgreSQL Docker GCP Terraform

Freelance

ML & AI Engineer

Jan 2018 – Present Remote
  • I build end-to-end ML pipelines: data cleaning, feature engineering, model training, deployment
  • I work directly with clients. No account managers, no translation layers. Just me, the data, and the problem
  • Recent work: demand forecasting for a retailer, document classification for a legal team, LLM-powered text extraction from PDFs
Python Pandas Scikit-learn TensorFlow OpenAI APIs SQL Docker
02

Projects

GedPi

Open Source

A batteries-included Pi package with an always-on workflow for clarifying specs, planning in bounded slices, and implementing with built-in verification. Ships repo-map awareness, skill discovery, and web search out of the box.

TypeScript Node.js CLI Agent Workflows
Learn more → GitHub ↗

GedCode

Open Source

A workflow layer on top of OpenCode that adds durable project memory, a plan-before-edit guard, TDD discipline, and skill discovery. Sessions stop forgetting, and the agent stops charging ahead without a plan.

TypeScript Node.js CLI Agent Workflows
Learn more → GitHub ↗

Ghost in the Web ↗

Open Source

A Chrome/Edge side-panel AI assistant that does things on the web with you, not instead of you. Scrapes pages, fills forms, chains across tabs, and exports to markdown, sheets, or PDF. Bring your own API key, everything stays on your machine.

TypeScript Chrome Extension AI Agents Web Scraping
03

Research

Reinforcement Learning vs Tree Search in 3D Tic-Tac-Toe

BSc Thesis

Lancaster University Leipzig · First-class honours

Compared tabular Q-learning, DQN, and Dueling DQN against Monte Carlo Tree Search in a 4×4×4 Tic-Tac-Toe environment. The study examined how each approach scales with board complexity, where neural function approximation outperforms classical methods, and where MCTS search depth still wins.

Python PyTorch Reinforcement Learning DQN MCTS
04

Skills

AI & ML

LLM system design Prompt engineering OpenAI APIs Anthropic API TensorFlow Scikit-learn PyTorch

Programming

Python TypeScript Rust Go Java C Lua

Infrastructure

Docker PostgreSQL Linux Git GCP Terraform CI/CD
Microsoft Certified: Azure AI Fundamentals (AI-900)
Anthropic API Certification
05

About

I started freelancing in 2018 because I wanted to work on real problems with real data. Since then I've built everything from demand forecasting models to LLM-powered document processors. I work directly with clients, no middle layers, which means I have to understand the actual business problem, not just the technical spec.

In 2025 I finished my CS degree at Lancaster University Leipzig (first class). Now I'm at Kibo AI, where I build LLM-powered tools that need to actually work, not just demo well.

Outside of work I build open-source developer tools (GedPi and GedCode are the current ones), run a homelab where I self-host everything I can get away with, and occasionally pick up an Arduino project of questionable ambition. I also bike and lift weights.

German Native · English C2 · Romanian Fluent
06

Contact

ed.gyarmati@proton.me → github.com/edgyarmati → linkedin.com/in/eduard-david-gyarmati → CV →