Building high performance, AI-native engineering teams
summary
Vercel rearchitects for AI by treating evals as automated tests, using Brain Trust for a test harness, and storing model configs in feature flags at the edge with ~25ms reads. A VZ prototype generates production React code from prompts, turborepo plus Next.js keeps builds fast at ~2 minutes, and an AI SDK called aisd lets teams swap providers and prompts instantly so a fine‑tune can go to production in ~4 seconds. Everything flows into Vercel Analytics for tight feedback.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:02AI-native engineering teams
- 3:09Software 1.0 to 2.0
- 5:34AI-native flywheel framework
- 6:58Evaluations foundation
- 9:45Data and fine-tuning
- 10:50Unified AI SDK
- 12:17Software 1.0 tools
- 13:40Preview environments
- 14:44Feature flags
- 17:54Scaling AI apps
- 19:00Instrumentation
- 21:11Measuring AI systems
- 22:35Antifragile systems
related talks
46mBecoming an AI engineering company
50mOpening Keynote
30mAccelerate AI to prod
18mAutomating 90% of support tickets: How we built Vercel's support agent
19mAutomating 90% of support tickets: How we built Vercel's support agent
09mAI for React Developers: Opportunities, Learning, and Innovation
25mBuilding Full Stack AI Applications with
23mBecoming an AI engineering company