What to build
A TypeScript application that provides a real-time agent assist experience — streaming a live phone call or conversation through Deepgram STT with speaker diarization, while an LLM generates suggested responses and surfaces relevant information based on the ongoing transcript.
Why this matters
Contact center agent assist is one of the highest-value enterprise use cases for real-time speech-to-text. Support agents handling live calls need a sidebar that shows the live transcript with speaker labels, highlights key entities (account numbers, product names, issue categories), and provides AI-suggested responses or knowledge base articles in real time. Developers building this pattern need to see how to wire Deepgram's streaming STT with diarization into an LLM for real-time response generation — a common architecture that currently has no reference implementation.
Suggested scope
- Framework: Express.js backend + React frontend
- Deepgram APIs: Streaming STT (Nova-3) with diarization and keywords
- Key features:
- Audio input from microphone (simulating a call)
- Real-time transcript with speaker labels (Agent vs. Customer)
- Entity highlighting in the transcript (names, numbers, products)
- AI response suggestions panel powered by an LLM (OpenAI or Anthropic)
- Knowledge base lookup triggered by detected topics
- Conversation summary that updates as the call progresses
- Complexity: Medium — real-time streaming with parallel LLM processing
Acceptance criteria
Raised by the DX intelligence system.
What to build
A TypeScript application that provides a real-time agent assist experience — streaming a live phone call or conversation through Deepgram STT with speaker diarization, while an LLM generates suggested responses and surfaces relevant information based on the ongoing transcript.
Why this matters
Contact center agent assist is one of the highest-value enterprise use cases for real-time speech-to-text. Support agents handling live calls need a sidebar that shows the live transcript with speaker labels, highlights key entities (account numbers, product names, issue categories), and provides AI-suggested responses or knowledge base articles in real time. Developers building this pattern need to see how to wire Deepgram's streaming STT with diarization into an LLM for real-time response generation — a common architecture that currently has no reference implementation.
Suggested scope
Acceptance criteria
Raised by the DX intelligence system.