The NLU Orchestra
summary
Claims that linear pipelines fail for real conversations and replaces them with a DAG of micro-models that each expose a small, standardized output contract so components mix and match. Implements an orchestrator/conductor that runs nodes by topological order, supports A/B tests and small regex models alongside transformers, and debugs errors more granularly. Runs on booking.com chat (Facebook Messenger) and reports 12–15 models in production, improving deployability and experimentability.
AI-written summary · about
chapters (13)
- 0:00NLU Orchestra intro
- 1:59Linear NLU pipeline problems
- 8:44Conversations are nonlinear
- 10:08From pipelines to graphs
- 12:16Standardizing component contracts
- 15:53Orchestrator role
- 16:57Micro-model architecture benefits
- 18:41Model reusability and flexibility
- 21:53Subgraph experiments
- 23:16Proposed solution summary
- 24:00Production deployment and scale
- 25:04End-to-end approaches comparison
- 27:11Conversation state tracking
related talks
45mArchitecture of an NLP Deployment
30mBridging Classic ML Pipelines with the World of LLMs
33mBuilding Resilient (ML) Pipelines for MLOps
51mOrchestrating Large Language Models (LLMs) with CrewAI
01h 04mBuilding Expertise Into Generated Conversations
48mUnlocking LLM Potential: From Exploration to Integration
11mScaling Conversational AI in Production
38mNatural Language Understanding at Scale