An AI connector for a community of 400 SME owners, working in 8 days

The founder runs a community of Singapore SME owners, about 400 today with 1,000 as the goal. Most of the value of a community like that is who meets whom, and that depended on one person: the founder, remembering who does what and making the introductions himself. It did not scale past his own memory and his own evenings.

He had seen Boardy, an AI that talks to people, learns what they need and introduces them to someone who can help. He wanted the same thing for his members, inside a paid community, under his own name. I asked how the business makes money and what the goal was before writing any code. Eight days after his first message, he had a working version in his hands.

Stage	What it does	Status
1Study what works	Takes apart how Boardy onboards, suggests and introduces people, and makes that the spec	Done
2Get to know each member	Onboards each member in a private chat, by text or voice note	Beta
3Find the match	Searches the member catalogue and says why two people should meet	Beta
4Make the introduction	Asks both sides first and shares contact details only when both say yes	Beta
5Keep it safe	Checks access, caps spend and was attacked by two review agents before launch	Beta

The spec was not invented. Boardy already shows what good looks like, so the research took apart how it onboards people, how it suggests a match, how one introduction is asked at a time, and how fast it replies. The rule from the start was to copy what works and not add features Boardy does not have. To check the result, fifteen real Boardy messages were compared against 49 of the assistant's own messages in beta, and three gaps were closed.

A member starts a private chat on Telegram. The assistant asks a few questions about who they are, what their company does, who their customers are and who they want to meet, and builds a short profile from the answers. Members can answer by voice note, which is transcribed first. Every profile change goes through a save screen before it is stored.

The member profiles form one catalogue. When a member asks for someone, the assistant searches that catalogue and explains in plain words why a particular person is worth meeting. Matching runs on Gemini with Claude as the fallback, and chat runs on Claude. Searches are capped per member per day, so the cost of matching stays predictable.

There is one live ask at a time, the way Boardy does it. The assistant asks the other member whether they want the introduction, and contact details are shared only when both sides say yes. At the moment of the yes it checks again that both people still have access and still consent, so an introduction cannot complete for someone whose membership has just ended.

She is built for paying members and event guests, with a gate that checks access. Each member has a daily spend cap, chats are private only, and forwarded messages are marked as third-party text so they are never trusted as the member's own words. Before launch, two adversarial review agents attacked the code without seeing its defences. They found eight real issues, including one that would have let a single member spend about 30 times their cap, and all eight were fixed and tested.

• A working version 8 days after the first message, demoed to the founder on 23 September.
• In beta since 25 September, with real people talking to her before the stage launch.
• About 1,355 automated tests across the service.
• Measured capacity of about 240 onboardings and 30 replies a minute, sized for a room full of event guests.
• A fixed running cost of about S$42 a month at current vendor prices.

The founder's words after the first version: "The guy's a magician." The launch is on stage at the founder's summit on 15 October 2026, and this page will carry what she does there once it has happened.