Dmitry Petrov

Dmitry Petrov is the co-founder and chief executive of DataChain and the creator of DVC (Data Version Control), the open-source project that brought software-style versioning to machine-learning datasets and experiments. He now builds infrastructure that helps AI agents interpret video, sensor readings, medical imagery, and other complex physical-world data without continually repeating expensive computations.

From reproducible machine learning to physical data

Petrov holds a doctorate in computer science and previously worked as a data scientist on Microsoft’s Bing team and as a software engineer at Node.io. At Microsoft, he confronted a basic mismatch: developers could track changes to code, but the datasets, models, and experiments underpinning machine-learning systems were harder to reproduce. He created DVC to extend Git-centered workflows to large datasets and pipelines, then co-founded Iterative with Ivan Shcheklein to build on that open-source foundation.

In July 2024, Petrov introduced DataChain, an open-source system for preparing and curating unstructured datasets with machine-learning models and language-model APIs. When lakeFS acquired DVC in November 2025, he did not join the acquiring company; he and Shcheklein continued independently with DataChain.

The infrastructure behind useful data agents

• Typed context for unstructured data. A small video collection can contain vast hierarchies of frames, detected objects, labels, and relationships. Petrov’s open-source DataChain project turns these hidden structures into persistent, queryable datasets using Pydantic schemas and Python-to-SQL workflows.

• Native media, structured metadata. Petrov argues that video, images, and audio should remain in their native formats, with searchable metadata stored separately and linked through file references, versions, and timestamps.

• Agent harnesses for physical-world data. His dashcam-video demonstration applies a YOLO object-detection model once, then answers subsequent questions from stored detections. The architecture combines typed schemas, parallel execution, checkpoint recovery, incremental processing, and dataset lineage connecting source files, processing code, and results.

Petrov has also applied these ideas to neuroscience and medical-imaging datasets, where researchers repeatedly revisit EEG recordings and MRI scans. His focus has expanded from preserving data history to making complex data computationally accessible and reusable by both researchers and AI agents.