Anton Troynikov

Anton Troynikov is co-founder and chief executive of American Terawatt, which develops high-voltage direct-current transmission infrastructure for industrial electricity users, including AI data centers. He previously co-founded Chroma, the open-source AI application database, tackling two essential constraints on AI: useful information and reliable power.

Born in Ukraine and raised in Australia, Troynikov studied computer engineering and economics at the University of Melbourne, pursued graduate robotics studies at the Technical University of Munich, interned at Google, and co-founded Geddit in Berlin. He subsequently worked in robotics and computer vision, including vehicle-camera mapping, collaborative-robot perception, and engineering at Iris Automation.

With Jeff Huber, he founded Chroma to give AI developers an open-source embedding database capable of storing, retrieving, and updating application-specific information. His earlier Stable Attribution project explored identifying training images that influenced generated pictures, potentially enabling creator attribution and compensation; missing information about the artists behind training-image URLs remained a fundamental obstacle. He explored those challenges in an interview on generative-art attribution.

• Agent memory must update itself. Troynikov advocates retrieval systems that incorporate human feedback, an agent’s observations, and previously acquired skills. NVIDIA’s independently developed Voyager Minecraft agent illustrated the approach using Chroma as retrieval infrastructure.

• Nearest neighbors are not necessarily relevant. A database containing only bird documents can still return superficially similar results for a question about fish. Troynikov argues for retrieval abstention, metadata filtering, keyword signals, re-ranking, and task-specific relevance judgments to prevent irrelevant context from degrading model performance.

• Application-specific retrieval evaluation beats generic benchmarks. Embedding models should be tested against real production data and human relevance judgments. His AI Engineer Summit presentation also explored document-aware chunking and detecting semantic boundaries through token-prediction perplexity or discontinuities between embeddings.

• AI infrastructure ultimately depends on electricity. At American Terawatt, Troynikov is developing high-voltage direct-current transmission for large industrial loads, including data centers, addressing the physical power-delivery constraints beneath large-scale computing.