What we built
I was the technical lead on an AI chatbot for Roku’s customer support, covering. The system plugs into multiple LLM providers (Claude V2, GPT-3/4) through AWS Connect to handle customer queries with natural language understanding.
Results
Per the vendor’s published case study of the deployment:
- 48% AI containment, resolved without a human agent
- Chat share of support contacts rose from 30% to 43%
- Phone share fell from 70% to 57%
- Response time went from hours (human queue) to seconds
How it works
The platform pairs AWS Connect for comms infrastructure with a custom AI orchestration layer in React and Node.js:
- Multi-LLM routing picks the right model per query (Claude for nuanced reasoning, GPT for speed)
- RAG pipeline pulls from Roku’s knowledge base so answers stay accurate and contextual
- Escalation to humans happens automatically for complex issues, with full conversation context passed along
- Feedback loop uses interaction data to keep improving resolution accuracy
Technical details
- React chat widget on the frontend, Node.js orchestration layer, AWS Connect for the communication backbone
- Conversation context management keeps multi-turn dialogues coherent
- Fallback strategies make sure no query goes unanswered
- Monitoring dashboards track resolution rates, escalation patterns, and per-model performance
In the press
Roku’s AI transformation using the platform we built was covered by Quiq: Roku’s Agentic AI CX Transformation.
