Getting 150 builders from the lobby to Room 209 without a queue

AI Tinkerers Doha meets in the Tech 2 building at Qatar Science and Technology Park. For Round 4, on 19 October 2026, the organisers approved about 150 builders for Room 209 on floor 2. Most had never been in the building, and most would arrive in the same half hour, while one or two volunteers ran the desk.

I built Ada for that half hour and gave the talk about her at the event. She's a voice concierge on a lobby screen: she gives directions, checks guests in by QR pass or by name, and names people worth meeting. The model talks and picks tools, and plain Python rules decide who gets in. In 28 scenes spoken over crowd noise, nobody was admitted who shouldn't have been.

28

end-to-end scenes, crowd noise mixed into the audio

0

wrong guests admitted

0 to 1

cloud calls per check-in

Watching one guest's evening

A talk about a kiosk is easy to fake with slides. A slide can't show what the kiosk heard in a noisy lobby, or what it got wrong.

I opened the deck on a recording instead: one guest, start to finish, on the real Godot kiosk and server with lobby chatter in the microphone. The captions show what Whisper heard, slips included. Pick a step to jump to it.

Starting from the research

A meetup lobby has problems a language model can't see: a rush at the door, strangers who won't say their name out loud, and a room too loud to hear in. I had no time for interviews before the event, so the design started from published research and the event's own data.

I read check-in benchmarks, wayfinding studies, work on talking to strangers, and Whisper's own noise results. Each finding turned into one rule for Ada.

1. Most guests arrive in the first half hour.

   Check-in vendors plan for 60 to 70% of attendees arriving within 30 minutes of doors opening, and aim for under 60 seconds per person. Self-service kiosks see 15 to 30% of first-time users needing help.

   So: Ada has to clear the 6:00 to 6:30 rush on her own and send only the exceptions to the desk.

   Nunify, Event check-in time benchmarks (vendor data, 2026) ↗

2. People get lost inside buildings too.

   Wayfinding research puts the cause in three places: the building's layout, the signs, and the visitor's own picture of the space. First-time visitors have no picture yet.

   So: Directions name a landmark and a count ("the lifts to the right of reception", "the third door"), and a second kiosk on floor 2 picks the visitor up.

   Carlson, Hölscher, Shipley & Dalton, Getting Lost in Buildings, Current Directions in Psychological Science (2010) ↗

3. People hold back from talking to strangers, and misjudge how it goes.

   After a first conversation, people consistently underestimate how much the other person liked them. Separately, professional networking for instrumental reasons makes many people feel uncomfortable.

   So: Ada can make the first move for a guest: name someone who works on their topic and say where to find them.

   Boothby, Cooney, Sandstrom & Clark, The Liking Gap in Conversations, Psychological Science (2018); Casciaro, Gino & Kouchaki, Administrative Science Quarterly (2014) ↗

4. Speaking to a device in public feels exposed.

   In studies of voice assistants, people were less willing to say private information out loud in public places than at home.

   So: A guest never has to say their name: holding up the QR pass is the first option, and Ada never reads names or the guest list aloud.

   Easwara Moorthy & Vu, Privacy Concerns for Use of Voice Activated Personal Assistant in the Public Space, IJHCI (2014) ↗

5. Speech recognition gets worse fast as a room gets louder.

   Whisper's own paper shows every model it tested degrading as background noise rises, with white noise and with pub noise. People also raise their voice and change how they speak in noise (the Lombard effect).

   So: Push-to-talk marks exactly when the guest is speaking, the reply is captioned on screen, and anything with a right answer (check-in) goes through rules, not recognised words.

   Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (2022); Zollinger & Brumm, The Lombard effect, Current Biology (2011) ↗

6. Doha is an English-speaking room with many first languages.

   About 88% of Qatar's residents are foreign nationals. Arabic is the official language and English is the language of business.

   So: Ada speaks English by default and handles accented speech and names from many languages. A name has to match one guest clearly, so a mishearing goes to the desk instead of admitting the wrong person.

   Demographics of Qatar, Wikipedia (read 19 Oct 2026) ↗

Framing the problem for four people

One kiosk serves four people with different stakes. A first-time guest wants the room and no queue. A presenter wants to set up early and be found afterwards. The volunteer at the desk wants only the hard cases. The organisers want an accurate headcount and a guest list nobody ever reads out.

I turned those stakes into three questions to design for:

• How might we get a guest from the front door to Room 209 without asking a person?
• How might we check a guest in within seconds, in a loud lobby, without them saying their name out loud?
• How might we help a guest start one useful conversation without exposing anyone who didn't opt in?

I drew the personas from the research and the event's data, with no interviews yet, so they are proto-personas and every guest is invented. The four edge cases at the bottom shaped the check-in rules more than the main four did.

Mapping Nadia's evening

Nadia is the primary persona: a speech researcher new to Doha, with her QR pass on her phone. Her evening runs from two days before the event to leaving through the gates at 9:10, and it dips twice: arriving in a loud lobby, and walking up to strangers.

I drew the map before building anything. Each dip became a feature. The room number goes on screen because the lobby is loud, and Ada names people who opted in because Nadia won't walk up to random ones.

Storyboarding four moments, then breaking them

The storyboard has four panels: the lobby, the pass scan, the second kiosk on floor 2, and networking. Each one looked simple on paper and stopped being simple once I listed what could go wrong in it.

I gave every panel its edge cases, and each edge case became a test scene. The stills come from those test runs, so they show the lobby kiosk: the floor 2 kiosk is still only a panel in the storyboard.

Keeping admission out of the model

A language model that can take actions will sometimes take the wrong one. OWASP lists this as Excessive Agency ↗. At a door, the wrong action means admitting someone who isn't on the list, or telling a stranger who else is here.

A first name alone gets the same reply whether one guest or two share it, so the answer gives nothing away.

I split the job in two. The model understands the guest and picks a tool. Rules in server/meetup.py decide every check-in and write the sentence Ada says, so the model never sees a full name or a pass code.

The rules can return 16 outcomes, and every one ends somewhere a guest can act on: admitted, asked for more, sent to the desk, refused, or asked to try again. Filter them by what happens next.

Names are the tricky part, because Whisper spells them however it likes. The server compares how names sound, so “Kabib Bi Han” can match “Cabibihan”, but only when one guest is the clear winner: at least 0.72 alike, and 0.15 ahead of the next closest guest. A first name alone gets the same reply whether one guest or two share it, so the answer gives nothing away.

Tracing each flow to the function

Two programs share one Mac. Godot draws the lobby and owns Ada's body. A Python server built on Pipecat listens, thinks and speaks. They talk over one local WebSocket, and only the text of each turn leaves the machine.

The stack is chosen so nothing personal leaves the building.

I named every box after the real function, so the talk could follow one question from the microphone to Ada's footsteps. Pick a flow and step through it.

The stack is chosen so nothing personal leaves the building. Whisper and Kokoro run on the Mac's GPU, OpenCV reads the pass locally, and the model sees only the words of the turn and the guide entries.

Inside Godot, one frame runs in a fixed order, with audio on its own thread beside it. That order is why a guest can cut Ada off and the walk to the lifts keeps going.

Testing 28 scenes in a loud room

Unit tests can't tell you whether a guest gets through the door. So every scene from the storyboard was spoken through the whole stack, Godot to server and back, over synthetic lobby chatter from ten macOS voices.

A mishearing makes Ada ask again. It can't admit anyone, because admission never trusts the transcript.

When a loud room got in the way, it was the hearing: Whisper turned “AI Tinkerers” into “AIVP” in a busy lobby, a guest cut in as a check-in landed, and the loudest room drowned the question. Ada asked again each time, and none of it let in the wrong guest.

Two of the scenes are on video at the top of this section. Scene 27 is the loudest room; scene 23 is a guest cutting Ada off mid-answer.

Pick a scene to see what the guest said, what Whisper heard, what Ada said back, and how loud the room was.

Holding to six principles

A kiosk at a door makes a small call for every guest, and a guest notices each one that goes wrong: a name read out to the room, a wrong admission, a reply lost in the noise, a refusal with nowhere to go next.

I held every decision above to six principles: one for each of those four, one for knowing when a guest has finished talking, and one for keeping personal data on the Mac. Each principle has a measure the test run reports.

Planning the next round

The demo runs one kiosk on one Mac with invented guests. Before it meets a real door, it needs the second floor, a better pass reader, and people who aren't me.

I split what's left into two lists. The build list comes from the rehearsal logs, where OpenCV missed passes on bright phone screens and stray words from the room started replies. The research list checks my proto-personas against real guests, starting with a count at the desk from 6:00 to 6:30 at the next event.

Notes

• Meridian Tech Park, its rooms and every guest in the demo are fictional.
• Conversation state lives in memory, so a restart forgets the chat but not the check-ins.
• The talk site ↗ has the 16 slides this post is built from, and replays one recorded evening with no server.

Sources

1. LLM06: Excessive Agency ↗, OWASP Top 10 for LLM Applications, 2025.
2. Pipecat ↗, the open-source voice pipeline the server is built on.
3. Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision ↗ (Whisper), 2022.
4. The research findings above link their own sources.