Skip to content
Pedro Gomes
All work

Sofia

Sofia is the voice agent Gomio uses to speak with people considering its services. She answers questions about the work, gathers the contact details and need behind each inquiry, then sends the lead to the team. A separate decision console shows how the agent applies Gomio’s rules.

Talk with Sofia (opens in a new tab)
Sofia’s console after a request to reprogram her: the reply on the left and, on the right, the decision card with the rule code INJECTION_BLOCKED

Who has this problem

A prospective client often starts with a broad question: what does this company build, and can it help with my process? Sofia handles that first conversation for Gomio, explains the services within their stated scope, and collects enough context for a person to follow up. The system also needs clear boundaries around uncertain project pricing, unsupported requests and attempts to change its instructions.

My role

Built for Gomio: the business configuration, voice runtime integration, policy layer, lead capture and delivery, decision console, and deployment.

How to try it

  1. Talk with Sofia in the browser and ask what Gomio does or describe the system you need to connect.
  2. Share a name and contact number when you want the team to follow up. The agent records the inquiry and sends it to Gomio.
  3. Ask for a fixed quote for a custom project or try to override her instructions, then inspect the decision shown beside the conversation.

Usage conditions

  • Live conversations last at most 3 minutes, with a warning before the cut.
  • 2 conversations at a time, 30 per day across everyone, and 3 per hour per connection.
  • A kill switch stops new conversations without a redeploy.
  • The ready-made cases make no speech or model calls.

Public access has usage limits to keep AI service costs under control.

Decisions I made

  • The voice session streams speech through transcription, the configured language model and speech synthesis. The model carries the conversation and proposes the name, phone and reason; the runtime validates those fields and controls when it asks for each one.
  • Gomio’s service catalogue, FAQs, refusal rules and lead fields live in a business configuration. A deterministic policy layer handles defined boundaries, including unsupported unofficial integrations and project quotes that need a technical scope. The language model still interprets ordinary questions and drafts their replies.
  • When the required contact fields are complete, the lead workflow sends the inquiry to Gomio’s contact inbox. Sofia does not book meetings or write into a prospect’s business systems.
  • The decision console receives structured events from the voice session. It shows the rule code, decision and outcome; the public decision record stores a fingerprint of the caller’s words instead of the raw utterance.
  • Ready-made cases replay evaluation traces without making a new speech or model call. They let a visitor inspect representative decisions alongside a live conversation.
  • The worker enforces limits on call duration and concurrent sessions, with a shared daily cap and a kill switch outside the conversation. This contains service cost without asking the model to enforce its own limits.

Limits that remain

  • The public conversation runs in the browser. Phone-number routing and a human call transfer are not demonstrated here.
  • Sofia captures and emails an inquiry; she does not book a meeting or connect to a prospect’s CRM or calendar.