AUZ Case Study · AI Automation
An always-on assistant that understands, guides, and qualifies.
Vivi is the intelligent assistant Auztec built for Vision Touch Ltd’s website. It understands natural-language, typo-tolerant questions, recognises 140+ Greater London areas and postcodes, answers from the site’s own content, and qualifies enquiries through a guided flow — all without a mandatory paid AI service, with a one-switch upgrade to Cloudflare Workers AI.
- Client
- Vision Touch Ltd
- Sector
- Conversational AI · Lead Qualification
- Category
- AI Automation
Vivi
AI Automation
Qualify enquiries 24/7 — without a fragile backend
A construction firm’s website visitors ask wide-ranging questions — services, areas, timelines, quotes — often with typos and local place names. The brief called for an assistant that could answer instantly and offline-capably, qualify real leads, and run at near-zero cost on serverless hosting, with a clean upgrade path to a hosted LLM when the business wanted it.
A three-layer engine: local intent, optional AI, safe fallback
We engineered a small conversational engine that scores keywords over a knowledge base built from the site’s own content, with bounded fuzzy matching and light stemming for typo tolerance, plus conversation context that remembers the current service and area. An optional server-side AI layer (Cloudflare Workers AI or any OpenAI-compatible provider) is consulted only for low-confidence questions, and a rule-based fallback handles human handover, lead capture, and email drafting.
Capability highlights
The engineering and design work behind the product.
Typo-tolerant understanding
Bounded Levenshtein fuzzy matching and light stemming resolve “rennovation”, “kitchin”, “illford”, and more.
140+ areas & postcodes
Recognises all Greater London districts plus postcode-prefix detection (IG1, HA1, SE15 …).
Conversation context
Remembers the current service and area, so follow-up questions and a plain “yes” continue the right thread.
Guided lead qualification
A nine-step flow captures the enquiry with project type and location pre-filled, then drafts an email.
Knowledge-grounded answers
A retrieval fallback fuzzy-searches the whole knowledge base before ever saying “I’m not sure”.
Privacy & safety guardrails
No server-side conversation storage, input caps, rate limiting, and grounded prompts — no prices, no promises, no invented facts.
What the build ships with
Understands natural-language, typo-tolerant questions
Recognises 140+ Greater London areas plus postcode prefixes
Nine-step lead-qualification flow with pre-filled project type and location
Answers grounded in the site’s own knowledge base — no mandatory paid AI
37 automated intent checks in the test suite
One-switch upgrade path to Cloudflare Workers AI
Multidisciplinary delivery, one team
The stack
- Vanilla JavaScript widget (lazy-loaded)
- Shared local intent engine
- Cloudflare Pages Functions (/api/chat, /api/chat/lead)
- Optional Cloudflare Workers AI / OpenAI-compatible layer
- Web3Forms lead pipeline
- Automated intent test suite
Related work
Have a project like this in mind?
Tell us what you're building. We'll help you turn it into a scalable digital product.