Eleven AI Workstreams, One Architecture: What Consolidation Actually Took
About this session
AI programs rarely fail because a team cannot build a prototype. They fail because too many teams build overlapping systems without shared evidence, ownership, or launch criteria. In my final role at TikTok, I led AI strategy for global Trust & Safety and worked with 27 data scientists across three regions. We converged 11 separate LLM workstreams into one architecture serving 17 markets. I will explain what made consolidation necessary, where the work stalled, and the decision gates we used to separate promising demos from systems ready for production. Attendees will see an illustrative before-and-after map. The metric that moved most was architecture fragmentation: 11 workstreams to one architecture. The map applies the decision loop I teach at Stanford: value, evidence, accountability, release boundaries, monitoring, and stop rules. It is a teaching reconstruction, not a former-employer system diagram.
Speaker
Key takeaways
- Map overlapping AI work around shared user decisions, not team ownership.
- Define deployment-specific evidence and go/no-go criteria before production.
- Assign owners, review gates, monitoring, and stop rules so scaling remains reversible.