Veloquence

Problem

Most of a B2B pipeline starts on the website, and almost none of it gets a real conversation there. Inbound leads wait about 47 hours for a reply, half of a rep’s time goes to prospects who were never qualified, and most of the buying journey happens before anyone from sales is involved (industry figures Veloquence cites). Forms are slow and scripted chat widgets stop at the first unexpected question.

System

Content goes into a knowledge base by upload, website crawl or connected sources, from PDFs and decks to audio and video. A Curator agent tags it, flags anything outdated and points out questions the agents can’t answer yet. Every item has an access level (Public, Identified, Evaluated, Customer), and the agents only surface what each visitor is cleared to see, with every reference logged and every approval versioned. A Flow Builder decides what each agent asks, when it hands off and what content appears at each stage, while the agent still reasons in real time. Companies set the voice, tone, speed and avatar, and drop each agent onto their site, product or event pages with an embed code. The internals belong to Veloquence, so I don’t share them here; my open-source voice interviewer, pipecat-demo, shows how I build real-time voice.

Hard parts

• Owning everything as a team of one: data model, backend, web app, infrastructure and deploys.
• Agents that reason freely but only say what the company approved, with a trail for every claim.
• Real-time voice and avatar conversations that feel natural.
• Carrying context from one agent to the next, from first question to demo and CRM.

Result

Live at veloquence.ai: one plan with every agent and unlimited conversations, a 21-day free trial, and its first design partners onboard. Built and run by one engineer.