Maker: the open-source AI event concierge that knows who you should meet tonight

Every meetup starts with the same three questions. What's on tonight, and where? Which table should I join? And the one that actually matters: who in this room should I be talking to?

Maker answers all three. It is a live 3D AI host for community events, built on the Perxona Connect Kit by Angie Chen, Perxona's marketing director. She is not an engineer. She vibe-coded Maker as her own event companion for the AI Tinkerers Taipei Kickoff on October 8, 2026, hosted by PicCollage, with six demo speakers and a room full of topic tables. By her own count, in one night Maker checked in 47 builders at the door, made 65 introductions between people who should meet, and answered 70 questions, including exactly what pizza was being served. Then she put the whole thing on GitHub under an MIT license. There is a public demo you can open on your phone right now, no sign-up needed.

This is the first post in a short series on Connect Kit projects that are ready to become real products. Maker is first because it already behaves like one, and because it has already worked a real door.

Image 2: the three questions every guest asks

What Maker does

Guests scan a QR code at the door and Maker runs on their own phone. No app to install. They check in, then ask by voice or by typing, in English, 中文 or 日本語. It also runs as a kiosk at the entrance, on one laptop or a tablet pointed at it.

Three jobs, in order of how often guests ask:

"What's on, and where?" Maker answers only from the event's own information, down to the pizza order. If something is not in the event file, it says "ask the staff" instead of guessing.
"Which topic table should I join?" It suggests one or two tables and says why they fit what you are building.
"Who should I meet?" It recommends up to three attendees who opted in, from what they wrote when they registered, and shows their LinkedIn as a QR code on screen.

If the venue Wi-Fi drops, the kiosk turns into a schedule board with QR codes until the network comes back. Nobody has to touch it.

Why a 3D avatar, not a chatbot

At a crowded door, a face that greets you, gestures and keeps eye contact is the difference between an FAQ page and a host. The 3D avatar, its expressions, gestures and lip-sync are Perxona's, loaded from Perxona when the page opens. Maker's voice comes from ElevenLabs, which is optional. Everything else is Angie's.

For developers, this is the part of the Connect Kit Maker uses:

<sv-presenter> with initializeWithConnectKey and the publishable key
present() for text, and presentWithAudio() to lip-sync its own ElevenLabs audio
playMotion() for gestures from the avatar's motion list, muteAudio() for a muted start
PRESENTER_STATUS and ALL_PERFORMANCE_FINISHED events
The server-side Connect API for avatars, scenes, voices and credit

A few lessons from earlier Perxona events are baked in. Replies are capped at about 320 characters, because long text has caused errors. present() never throws, so its result is checked. A 429 from the AI service retries with backoff and then says "one moment" instead of freezing.

Image 3: how it works, from the QR at the door to an introduction

Built so the safe thing is the only thing

Introductions at events are a privacy question before they are a product question, and Maker treats them that way.

The AI only ever receives people who said yes to being introduced. Being on the registration list is not consent.
Once check-in starts, only guests who are actually in the room are recommended. No-shows are never suggested.
Email and phone columns are dropped on import. Email-looking and phone-looking text inside answers is blanked out.
No web search, no scraping. Maker knows only what each person wrote on the form.
Nothing visitors say is stored. The stats folder holds anonymous counts only.
Every paid call is capped per phone, per guest, per address, per hour and per day, so a public demo cannot run up a bill.

No secret key ever reaches a browser. The page gets only Perxona's publishable key, which can make the avatar speak and nothing else.

Run it for your own event in about 30 minutes

The demo runs in two minutes. Your own event takes about half an hour.

Install (2 min). Clone the repo, npm install, copy .env.example. You need Node 22+ and Chrome.
Keys (5 min). Two Connect keys from the Perxona Console, one Anthropic API key, a staff password.
Your event folder (15 min). Copy the template. event.yaml is everything the concierge knows: schedule, venue, demos, topic tables, FAQ. persona.md is the avatar's name, greeting and personality. Pick an avatar, scene and voice from your Perxona catalog.
The opt-in form (5 min). A Google Apps Script makes the consent form for you.
Rehearse (3 min). One command tests keys, credits, the voice, answers in all three languages, matching, prompt injection and the offline board.
Phones and the cloud (optional). A ready Render blueprint puts Maker on the web so guests can use it on their own phones.

Running cost: a Perxona account, with a free trial available, plus a few dollars of AI credit. A busy evening is about US$5.

Image 4: set-up in about 30 minutes

Where this can go

Maker is an open-source project built for one community's events. The shape of the product, though, is already clear, and it fits several groups who pay for exactly this.

Meetup and community organizers. At 30 guests, the host can introduce people by hand. At the Tinkerers kickoff, Maker made 65 introductions among 47 builders in one evening while the organizer was busy running the night. At 150 guests, nobody can do that by hand. A door host that knows the roster and makes good introductions is the job organizers want done and never have time for.

Conferences and trade shows. Business matchmaking is already a paid service at B2B fairs. Maker's opt-in matching with a LinkedIn QR on screen is a lightweight version an organizer can run at the entrance, and a sponsor can put their name on.

Agencies and system integrators, especially in Japan. Reception and concierge avatars are a familiar category there. Maker is one event folder per client, one persona per brand, three languages out of the box. The avatar, voice and gestures are already done. The agency sells the setup, the content and the day-of operation.

Venues, coworking spaces and incubators. Weekly events with a changing roster and the same setup. Maker's staff page already handles the Luma guest export, check-in and walk-ins.

What would take it from demo to product is all app-level work: a hosted version with no laptop in the loop, sync with Luma or Eventbrite, sponsor branding on the door screen, and a post-event follow-up that sends each guest the people they met. None of it touches the avatar. That part is done, and it is the part that makes guests walk up to the screen in the first place.