Summary
Dmitry Petrov, creator of Data Version Control (DVC) and founder of DataChain, addresses a category of agentic AI failures that most published research ignores: unstructured physical data. While Anthropic has reported only 21% agent accuracy on data projects and OpenAI has published a six-layer context framework for structured business data, Petrov argues both operate in a comparatively comfortable world of tables and SQL. Physical data — video recordings, sensor telemetry, robot logs — presents a fundamentally different problem.
The core challenge is scale explosion: 2,000 video files can contain millions of nested objects (clips, frames, bounding boxes, confidence scores, object classes), and the standard approaches of JSON sidecar files on S3 or a separate centralized database both introduce unacceptable complexity for researchers. DataChain’s solution is to use Pydantic data models as the schema layer, eliminating the SQL island problem and letting Python serve as both the data definition language and the processing language.
The talk includes a live demo using Claude Code to analyze dashcam motion data: installing DataChain via pip, defining a Pydantic schema that maps video objects to a queryable database, and running parallel distributed processing across files. The execution engine handles checkpointing so failed jobs can resume without reprocessing. Petrov also supports Cursor and other coding agents in addition to Claude Code. DataChain is open source, and the demo shows how an agent can answer analytical queries (“how many pedestrians appear?”) over terabytes of video data using nothing but standard Python and the DataChain harness.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Hands On Build
