DEVOURED
AI Agents to Make Sense of Data at OpenAI (45 minute video)
OpenAI's internal data analyst agent, Kepler, uses automated code crawling and scoped semantic memory to query over 600 petabytes of data.
What: Bonnie Xu describes how Kepler uses AST-normalized LLM grading, daily Codex-driven code analysis, and Model Context Protocol (MCP) to provide context-aware data analysis while enforcing existing data permissions.
Why it matters: This demonstrates a shift toward agentic systems that actively curate their own context via automated repository crawling rather than relying solely on static metadata.
Deep dive
- Kepler serves as an internal AI analyst querying over 600 petabytes of data across 70k datasets.
- The agent uses MCP for tool orchestration, allowing it to perform multi-step query refinement.
- Automated Codex jobs crawl code daily to infer table lineage, grain, and freshness.
- The system uses three-level scoped memory (user, team, global) to store corrections and preferences.
- Evaluations rely on AST-normalized SQL comparisons to ignore semantically equivalent but syntactically different queries.
- Data security is maintained by reusing existing authentication and permissions, rather than bypassing them.
- Chain-of-thought streaming is used to audit agent assumptions in real-time.
Decoder
- AST (Abstract Syntax Tree): A tree representation of the abstract syntactic structure of source code, used here to verify that two SQL queries produce the same result regardless of formatting.
- MCP (Model Context Protocol): An open standard for connecting AI assistants to data sources and development tools.
- RAG (Retrieval-Augmented Generation): A technique that retrieves relevant context from documents or databases to improve the accuracy of LLM responses.
Original article
Full article content is not available for inline reading.
Explore all 69 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi