Skip to content
Marcello Jardim
← All workCase study

2026

Support automation

A help centre, and a phone line that has read it.

Role
Automation & AI integration
Years
2026
Stack
Claude Code, Vapi, Claude, ElevenLabs, Twilio, GitHub, llms.txt

The Fluss+ help centre and an AI voice agent that answers from it. Articles people, search engines and LLMs all read properly, drafted with Claude Code; and a voice agent on the phone or the web that answers from the same pages.

The signal path

  1. Transport

    Two ways in for callers: a phone number on Twilio, or a call from the browser. For readers: index pages, sitemaps, robots.txt and llms.txt.

  2. Cloud

    Vapi runs the call, with Claude answering from a knowledge base that syncs from the help-centre articles on GitHub. Claude Code drafts new articles, and a pipeline recaptures screenshots when the app changes.

  3. Interface

    The articles, one consistent format for people and LLMs alike, and an ElevenLabs voice on the line.

Context

Fluss+ and Gemlink had about 20,000 users between them, and every how-do-I and why-won’t-it question ended up with the support team. There was nothing written down to point anyone to.

The problem

Write the answers down once, in a way that works for a customer at the gate, a support agent mid-call, a search engine and an LLM. Keep them true as the app changes, and answer phone calls from the same source.

What I built

  1. About fifteen articles by hand, on using Fluss+ and troubleshooting the app and its devices, to learn what a good one looks like.
  2. From those, one format and a precise way of wording, so LLMs read each article as cleanly as people do.
  3. Article drafting with Claude Code: give it a new subject and it drafts the article to that format.
  4. Screenshots that keep up: when the app’s front end changes, the affected screenshots are recaptured automatically.
  5. Discoverability for search engines and LLMs: index pages, sitemaps, robots.txt and llms.txt.
  6. A voice agent on Vapi, with Claude handling the caller’s request and a natural ElevenLabs voice.
  7. Two ways to call: a regular phone number through Twilio, or a web portal that calls from the browser.
  8. A knowledge base synced from the help-centre articles on GitHub, so a fix to an article reaches the agent too.

Outcome

The support team answers from the articles, customers find them on their own, and the voice agent answers calls from the same pages.

© 2026 Marcello Jardim

Designed and built by me. React, Tailwind and a few React Bits, set in Bricolage Grotesque and Martian Mono.