At Klipy, I designed and built an AI game generation service where coding agents turn chat prompts into playable HTML5 games. Each agent works in a persistent Git workspace, and games are automatically built, checked in a headless browser, and published to the Klipy platform.
I also shipped a DeepSeek-based agent runtime after prototyping several architectures, and built quality checks that test real game interactions and trigger bounded automatic repairs. On Vertex AI, I created tools for generating game art and animation assets.
At Epam Anywhere, I developed a chat-based legal matter search application using LangGraph, structured SQL queries, and hybrid document search. I also worked on an anomaly detection system for Novartis’s order-to-cash cycle, ranking risky transactional windows and presenting insights in Streamlit.
At Bank of Georgia, I worked on credit scoring, product recommendations, Georgian speech-to-text, forecasting, and chatbot projects. My independent research includes porting Orthrus speculative decoding into vLLM and developing a streaming, multi-speaker speech-to-text pipeline.
