At Dreamlab Technologies, I built an AI agent platform for three clients, including a national political party, and launched its public assistant across web, messaging, a voice kiosk, and a humanoid robot. It answered more than 18,000 questions from over 10,000 people, and I cut median reply time from 5.3 seconds to 3.3 seconds.
I built a compliance agent that reviewed 306,667 administrative penalty cases in its first week, checking each case against the version of the law in force at the time. I’m now building its successor, where agents turn a 1,078-article code into a knowledge graph and derive checks from it.
I also serve open models on our own GPUs, including at air-gapped sites. Previously at Victory Technology, I served Qwen3-32B with vLLM and trained a production BERT text classifier; in a project, I trained and published a Kazakh named entity recognition model.
