Key facts
| Sector | Customer support / AI |
|---|---|
| Type | AI help chat with human handoff and an agent inbox |
| Running on | OF.luxury, VIPListed, BOQQS, ereQa and ruxox |
| What ruxox delivered | Widget, answer engine, live knowledge base, handoff to WhatsApp and Telegram, web inbox, deployment |
| Status | Live since October 2026 |
The brief
Five brands, five audiences, one small team. Visitors to a concierge, a nightlife booking site, a jobs platform, a finance platform and a software studio ask very different questions, often outside working hours and not always in English. We wanted every one of them to get a correct answer straight away, from that site’s own facts, and to reach a real person quickly when the question was a booking, a quote or something about their account. We did not want five chat subscriptions, a server to look after, or a widget that slowed the sites down.
What we built
One chat service that every brand plugs into with a single script tag:
| Part | What it does |
|---|---|
| Answers | Replies 24/7 in 8 languages, only from the site’s own knowledge base, and says so when the answer is not there |
| Live knowledge base | Read live from each site, so a deploy that changes prices or policies reaches the bot within minutes |
| Visitor details | Name and email first, so the team can follow up if the visitor leaves |
| Handoff | Bookings, quotes and account questions go to a person: WhatsApp for the luxury brands, live chat for the others |
| Agent side | Live chat answered from Telegram (one topic per visitor) or from a web inbox with replies, private notes, saved replies and reports |
| Widget | About 10 KB, styled to match each brand, isolated in a Shadow DOM |
The knowledge base is the part that keeps the answers honest. Each site publishes a plain text file of its facts, and the chat reads it live. Nobody edits the bot separately: if the pricing page changes, the file changes in the same deploy, and the chat is answering from the new version within minutes.
How it is built
Answers come from Claude Haiku 5.5 through OpenRouter. The conversation logic runs in a Supabase Edge Function with Postgres for storage, and the handoff uses the Telegram Bot API, so there is no VPS to patch and no per-seat SaaS fee as the team grows. The prompt is arranged so the large, stable part (the knowledge base and the rules) is cached: about 98% of each prompt is served from cache, which brings a warm answer to roughly $0.0002 and 3–4 seconds.
The widget is small and polite. It loads deferred, sits in its own Shadow DOM so a site’s CSS cannot break it (or the other way round), causes zero layout shift, and had no effect on Core Web Vitals on any of the five sites.
Privacy is part of the design rather than a policy afterthought: chats are deleted after 12 months, and the bot never asks for card details.
What this shows
A useful support chat is mostly plumbing done carefully: a single source of truth for the facts, a fast route to a person, and running costs low enough that nobody has to think about them. The model is the easy part. Running it on five live brands with different audiences is how we know the plumbing holds.
Where this fits
This is an AI implementation build, and it is the kind of AI-powered customer support we scope for clients: grounded in your approved content, connected to the people who answer your customers, and priced as a fixed-scope project — see straightforward pricing for the tiers. If you want one for your own site, send us a brief and you get a written scope within 48 hours.