I like working across the whole system, from an early experiment to something people use.
From prototype to production: building scalable agent systems and leading the engineering team.
An AI agent that runs daily operations for tour operators, DMCs and hotels: communication, quoting, itineraries, CRM, outreach and task tracking, all from one inbox.
Trained Kona, a 12B Georgian LLM, with a data and training pipeline built from scratch. The models and datasets are now open source.
This work brought together the full language-model pipeline: collecting and preparing data, continued pretraining, supervised fine-tuning, and preference optimization. It is one of the projects I’m proudest of.
What happens when many agents work on the same problem, and how should they communicate?
How can agents keep useful context, follow instructions, and make progress over longer runs?
How can an agent improve its own harness and the software around it?
My earlier papers include work on EEG and emotion recognition. My current interests are focused on language models and agents.
Papers on Google ScholarAgent harnesses, multi-agent systems, long-running and self-improving agents. I’m most active on X.