Session

Blue Origin — Using Data to Build a Road to Space

Overview

ExperienceIn Person
TrackArtificial Intelligence & Agents
IndustryEnterprise Technology, Manufacturing
TechnologiesDatabricks SQL, Unity Catalog, Databricks Agents
Skill LevelIntermediate
DOWNLOAD SESSION SLIDES

At Blue Origin, Databricks is the central data lakehouse for design, manufacturing, and launch operations. Scaling to thousands of pipelines without scaling headcount required rethinking how we manage data in the first place. We will share how we use AI and automation to build pipelines and tables with no human intervention required and how we automated data quality and metadata management to make our data AI ready. This enables federated data teams across the company to use trustworthy data and AI to build a road to space!

 

We will walk through a multi-agent system built on this foundation to manage satellite telemetry. An orchestrator queries a knowledge graph of prior episodes and a vector store of resolution procedures, then routes tasks to specialist agents for anomaly classification, forecasting, resolution, and command safety validation. Each execution is captured, clustered, and distilled back into procedural memory, so the system improves from its own operational history.

Session Speakers

Clark Stevens

/Director Enterprise Data Management & AI
Blue Origin

Theodor Talvac

/Senior Manager Ground Software
Blue Origin

Full Summary

Inside Blue Origin: Why a Rocket Company Runs on Data and AI

Blue Origin reframes itself as a data company that builds rockets. Every design choice, factory step, test and mission outcome depends on turning petabytes of structured and unstructured information into fast, trustworthy decisions.

FAQ


Time series databases excel at fast queries over timestamped data, which the ground system already handles for short term operations. Long term learning requires joining telemetry with runbooks, anomaly episodes, models and fine tuning sets under one governed platform with lineage and auditability. That need pointed to a lakehouse.

It scores every data object on four factors, ownership, uniqueness, description quality and documentation. Assets above roughly 70 percent are prioritized for agent access, while those below about 40 percent trigger workflows that improve their metadata with GenAI drafts and human validation. Agents reach vetted data sooner, producing faster and more trustworthy answers.

Mint replaced thousands of hand built pipelines with a metadata driven, event based system. System owners publish schemas, pipelines spin up automatically, and data lands in S3 and the Databricks Platform with average latency under 10 minutes. A single ingestion serves the Databricks Platform, knowledge graphs and vector stores, and publishing is self service across the company.

The Databricks MCP server is the primary standardized interface for agents to query governed data. It serves more than 3,000 monthly users and over a million invocations, with 5.5 times the usage of the next most used MCP. User growth is 26 percent year to date, while invocations are growing 180 percent month over month.

Memory, not just agents. BRAIN maintains tiered memory that captures entities, events and procedural heuristics across a satellite's lifetime, then closes the loop by writing every validated investigation back to Delta Lake as a structured episode. That institutional memory retrains models in a medallion architecture, so each investigation improves the next.