Voice AI Platform for Sales & SupportAILive
Chatterbots.ai
Also known as VoiceIntel
Platform for building AI voice assistants that hold a real conversation, answer from your own content, qualify the caller, and book the meeting.
Result
Voice assistants that qualify, book, and hand off
The problem
A chat widget answers a question and stops there. The visitor who was ready to buy still has to find a form, fill it in, and wait. Meanwhile support teams answer the same twenty questions by voice all day, and every one of those calls is capacity a booked demo would have been worth more than.
What I built
Built voice-to-voice assistants rather than a text bot with speech bolted on, so the interaction has turn-taking instead of a transcript. Each assistant is grounded in a knowledge base built from the customer's own documents, website, and question and answer pairs, and is given the actions that make a conversation worth having: capture the lead, book the appointment, take the payment.
The outcome
A multi-tenant platform with white-label branding, conversation and conversion analytics, and plans from a single assistant up to unlimited, running as a commercial product.
How it works
- 01
The visitor speaks, and the assistant answers in voice, in the language it detects, across 50+ languages with no language picker to set.
- 02
Answers come from that tenant's knowledge base: uploaded documents, a crawled website, and explicit question and answer pairs for the things that must be answered one exact way.
- 03
When intent appears the assistant acts, capturing the lead, offering an appointment slot, or taking payment through Stripe inside the same conversation.
- 04
Animated avatars, built on D-ID, give the assistant a face where a face helps and are left off where it does not.
- 05
Every conversation is logged and scored, so the operator can see which questions get asked, where callers drop, and what converted.
The AI layer
A speech-in, speech-out loop over OpenAI models, with language detection on the inbound turn and the tenant's knowledge base retrieved before generation. Curated question and answer pairs sit ahead of retrieval, so pricing, policy, and compliance answers come back exact rather than paraphrased.
The engineering layer
The platform work is the harder half: tenant isolation so no assistant can read another tenant's knowledge base, white-label theming and domains, plan and quota enforcement, Stripe billing, calendar integration for booking, and conversation analytics. GDPR obligations shape data retention, which is a schema decision rather than a policy page.
Key technical decisions
Voice-to-voice, not text with a speaker on top.
Speech to text, then a text bot, then text to speech leaves the caller waiting through three round trips. Designing for voice turn-taking from the start is what separates a conversation from an IVR.
Exact answers beat retrieved ones for some questions.
Pricing and policy go through curated pairs rather than retrieval. A paraphrased price is a support ticket, and possibly a refund.
Multi-tenant from the first schema.
Retrofitting tenant isolation onto a single-tenant product is a rewrite. Every table carries the tenant from the start, so white-labelling is configuration rather than a fork.