The problem
Inbound website prospects had no guided way to explore products or book time with a rep, so promising top-of-funnel leads quietly leaked away.
What I built
- Built "Q" on OpenAI gpt-realtime over WebRTC (ephemeral 120-second session tokens) with Claude as the reasoning brain — a multi-model split where gpt-realtime owns sub-second speech and Claude owns reasoning.
- A
navigate_apptool drives the site (opportunities, products, map, dashboard) with source and product-fit filters, shared with a deterministic text parser. - A
consult_qtool routes any product question to Claude over the product KB and returns on-screen content (cards, comparisons, lessons) that Q narrates step by step. - A
show_meeting_formtool captures leads → a durable JSONL record + an AWS SES email to the sales team; a lip-synced mascot is driven by real-time audio amplitude.
System architecture
🎙️
Architecture diagram — coming soon
gpt-realtime (voice) ⇄ tools (navigate / consult_q / meeting) ⇄ Claude brain ⇄ AWS SES. (Upload diagram here.)
Demo
▶
Walkthrough video — coming soon
A live voice conversation: navigate → ask about a product → book a meeting.
Stack
OpenAI gpt-realtimeWebRTCgpt-4o-transcribeClaudeAWS SESReact
Outcome
Guided demos & qualified meetings Converted passive site visits into guided product demos and booked sales meetings.