Bilingual AI Sales Assistant & AI Lab

The problem

A sales assistant is tempted to promise. Prices, certifications, delivery dates and features a product does not have are exactly the sentences a visitor remembers — so they must never come from a model’s imagination.

What I built

• Answers grounded in the site’s own knowledge index, with weak matches dropped and a hand-off to a person when nothing relevant is found.
• A commercial claim gate in code: nine kinds of unsupportable claim — prices, certifications, delivery times, invented figures — are checked in both languages; an answer is regenerated once, otherwise withheld, and shown only after it passes.
• Prompt-injection defence by authority separation: tools and indexes are chosen on the server, and outside text is fenced as data.
• Live voice over Azure OpenAI Realtime that shares one conversation with text chat, with eight separate defences against the assistant answering its own echo.
• Leads captured, scored and summarised; a conversation left idle is wrapped up automatically, with a transcript link for the visitor and a notification for the team.
• AI lab pages: a rule-based AI-readiness assessment, a try-your-document demo that stores nothing, a side-by-side model comparison and a RAG-quality dashboard for administrators.