CADEMA Voice Agent
A Czech-speaking first line for inbound customer calls.
The business problem
Inbound enquiries depended on a single person. Missed calls meant lost opportunities, while repetitive delivery questions interrupted other work.
My contribution
My role covered workflow definition, architecture, direction of AI-assisted implementation, service integration, testing and deployment. Platform and library authorship belongs to the third parties credited below; additional collaborators are not identified in the public project notes.
Twilio supplies telephony; ElevenLabs supplies speech. These are third-party services, not platforms I authored.
Architecture & boundaries
Twilio and ElevenLabs connected to a constrained knowledge base and prompt/state-machine flow. Czech SSML pronunciation tuning and O2 conditional forwarding handle local language and human escalation.
- 01Incoming callTwilio / telephone forwarding
- 02Constrained answerKnown delivery information + Czech speech
- 03Human hand-offCustom quotes, uncertainty and exceptions
Labelled architecture overview, reconstructed from project documentation. This is not a production trace.
Inspect the workflow
Synthetic reference demo · no customer data, live calls, LLM requests or shop writes.
Caller / sample Czech inputKdy mi dorazí objednávka?
Agent / scripted responseRád pomohu s doručením. Nejprve potřebuji údaje k objednávce; nebudu odhadovat termín bez ověřených informací.
Routing decisionRequest verified order context; no invented delivery date.
This demonstrates routing intent. It does not evaluate speech recognition, voice quality, latency or real telephone forwarding.
Failure paths & design decisions
Czech place names were mispronounced.
The project notes describe an SSML pronunciation dictionary and iterative listening.
Evidence boundary: A real recording and pronunciation evaluation set are not published.
Delivery answers could sound confident without support.
Replace unconstrained answers with a bounded knowledge lookup; route unsupported requests to a person.
Evidence boundary: The synthetic demo below shows the routing decision, not speech recognition quality.
Custom quotes should not be improvised by the agent.
Use a human escalation branch through conditional telephone forwarding.
Evidence boundary: Reference tests check custom-quote and uncertain-answer routes.
Delivery record
- Status
- Described as deployed in the existing project record; no live phone endpoint is published here.
- Users & scope
- Czech-language inbound customer enquiries for CADEMA, with human escalation for custom quotes.
- Dates & usage
- Delivery dates and current call volume are not recorded in the public evidence.
Results & measurement gaps
Answers routine delivery questions, qualifies buyers and routes custom quotes or exceptions to a person. Iteration focused on real calls, pronunciation and reliable hand-offs.
No public call-volume, latency, task-completion, escalation-rate or missed-call comparison is available. The useful claim here is the workflow and its documented design decisions.