Back to DailyYouTubeOctober 16, 2026Duration:1:34:17

Uber's AI infra, from the inside

Uber's AI infra, from the inside

TPThe Pragmatic Engineer

Full transcript

English

Summary:Uber Distinguished Engineer Uday Medisetty walks Gergely Orosz through two dozen internal systems: one model gateway with five guard models per request, a context graph grounding every agent answer, and Minion producing 15% of production PRs with human review on each.

Watch on YouTube
Core points (3)

Core points (3)

  1. 1Every model request passes one gateway where five guard models enforce policy, billing and audit.
  2. 2A context graph with 100+ node types grounds every agent answer, raising quality while cutting cost.
  3. 3Background agents produce 15% of production PRs, yet humans review every single one.