Voice agents
- Status
- active
- Built at
- qBotica, for enterprise clients
- Stack
- Python, LiveKit Agents, Deepgram speech-to-text, OpenAI models and voices, FastAPI, Next.js, Salesforce
- Since
- 2025
Logistics calls
The one that's deployed takes inbound and outbound calls for a logistics company. During the call it creates, looks up and changes quotes, answers the common questions, and routes the caller to the right department when a person is needed.
Voice assistants in the browser
A second line of work runs in a web page instead of on the phone, on LiveKit. A salesperson talks to the agent and it adds and updates leads in Salesforce, finds a lead by name, and emails the deck or welcome packet. A compliance version for a purchasing group answers questions about suppliers, contracts and rebates, reads a contract PDF you drop into the call, and draws charts on the screen while it talks. It switches persona depending on who's asking: a finance VP, a supplier manager, a category manager. These were built as client pilots on sample data.
What the work actually is
Getting a model to talk is the easy part. Most of the work is making it safe to let it act (more in Your agent will say it sent the email):
- Read back before writing. Before the agent saves a lead or opens a case it reads a short summary back and waits for a yes.
- Never claim an action that didn't happen. Models will happily say "I've sent that email" without calling the tool. The prompts forbid it in so many words.
- Append, don't overwrite. Notes added by voice are date-stamped and added to the record, never replacing what a person wrote.
- Keep the conversation moving. The reply starts generating before the speaker has quite finished, and when a cough or background noise is mistaken for an interruption, the agent picks up where it left off.
- Say less out loud. Some things shouldn't be spoken on a call, like pricing tiers or another supplier's data. The agent offers to send those securely instead.