≈1.5 s
Median reply delay in platform benchmarks
Down from 5.7 seconds during development; not a production service guarantee.AI automation · Voice agents
From a conversation to the next action.
An in-house platform for voice agents that answer questions, capture leads and connect bookings and follow-ups to business tools.
3 channels
Website widget, browser and phone
Connected tools
Bookings, APIs and post-call workflows
YUME / YUME VOICE AGENT PLATFORM03 /
YUME VOICE PLATFORMTalk. Capture.
Follow through.
Follow through.
Website / PhoneVoice agentBooking / Lead / Follow-up
CONVERSATION → CONNECTED ACTIONProject typeYume in-house platform
Use caseCustomer enquiries, bookings and lead capture
Yume’s roleArchitecture, real-time audio, AI integration and automation
StackPython, LiveKit, OpenAI Realtime, Twilio, n8n, Supabase
01 /
The challenge
Start with how the work really happens.
A useful voice agent needs to do more than answer questions. It needs to let callers interrupt, respond promptly and pass the conversation into a booking, a lead record or another business workflow.
Building each agent as a separate application would repeat the same integration work. We built a shared platform to configure different agents and tools.
02 /
The response
Shape the system around the people using it.
Each agent has a configuration for its instructions, knowledge, voice and tools. A shared worker loads the appropriate configuration for each call.
Visitors can speak through a website widget or browser connection, and phone calls connect through Twilio. The browser and phone bridges feed into the same agent workflow.
Agents can use booking tools, HTTP requests, webhooks and MCP integrations. After a call, the platform stores transcripts, captured leads and latency figures, then sends them to n8n for delivery to tools such as Google Sheets.
An admin panel supports agent configuration and transcript review. We tuned voice activity detection and removed transcript pacing that blocked audio, measuring the changes with benchmarks and simulated calls.
01Configurable agents and knowledge
02Website voice widget
03Browser and telephone connections
04Caller interruptions
05Booking and API tools
06Transcripts and lead capture
07Post-call n8n workflows
03 /
Evaluation
Measure the conversation before promising an outcome.
The supplied project report records median reply delay falling from 5.7 seconds to approximately 1.5 seconds during development. Simulated phone calls measured approximately 1.9 seconds.
Validation included 21 unit tests, latency benchmarks and browser, WebSocket and phone simulators. These results describe the test setup; response times vary with the connection and deployment.
04 /
What changed
The platform brings voice conversations and follow-up actions into one configurable system. New agents can share the same calling infrastructure while using different instructions and tools.
In-house development evidence from the supplied project write-up. No client deployment, conversion uplift or staff-time saving is claimed.
Could routine calls and follow-ups use a connected workflow?