A thousand people showed up to a conference whose existence was an idea a few months earlier. Two people decided it should exist, and then it did.
If you stand close to an event like this while it’s being built, you stop believing the version of the world where things happen because companies or institutions decide they should. They happen because someone wakes up and starts moving. The rest is logistics.
The accessibility is what stays with me. There was no gatekeeper, no committee, no incumbent organization whose permission was required for AI Engineer Singapore to exist. Rachael wanted it to exist, said so out loud, and a city’s worth of builders showed up. The bar for “you can pull this off” is much lower than the bar most people assume. If you have structured opinions about why a thing should exist, and the patience to keep showing up to the boring parts, the rest is more available than you’d think.
That changes what you think you should be doing. If two people can decide on a Tuesday that a thousand AI builders should be in a room together, then most of the things on my own list that feel “too big” are not too big. They are unstaffed.
What I’d do differently next time is the operations layer underneath all of it. Most of the friction came from prep we hadn’t done before doors opened.
Speakers, bios, headshots, talk titles, session times, room assignments. Every one of those existed somewhere, but not in one place, and not in a structure that any downstream artifact could pull from. So every brochure, every Canva carousel, every Instagram caption became a small re-discovery project. Find the bio. Confirm the title. Match the headshot. Resize. Post. Repeat for the next session in two hours.
The fix is obvious in hindsight. One Notion database, speaker as the primary entity. Bio, headshot, talk title, abstract, session time, room, social handles, sponsor affiliation, all as fields. Canva pulls from it. The website pulls from it. The brochure exports from it. Every piece of content downstream becomes a view on the same row.
The second thing I’d build is a live-event LLM. Load it with the run of show before the event starts. Pipe transcripts in as sessions happen. Then during the day, ask it questions. What just happened in room two. What was the quote from the Carousell talk. Which speaker is up next and what should the caption mention. Live content during a conference is almost entirely about reducing latency between “something interesting was said” and “it’s on Instagram.” Most of that latency is human, and most of the human latency is recall. An LLM that has been listening the whole time fixes the recall part for free.
Both of these are pre-event work. The smoothness of an event during is mostly a function of how much structure you built before.
Justin Baird from Tesseract AI brought a woman with ALS onto the stage. They put an EEG headband on her, ran her neural signal into a robotic arm, and the arm painted. She was painting. The room understood what was happening a few seconds before the demo finished, and then everyone was crying at the same time, including me. There was nothing to say about it. The technology, the person, the gesture, the public-ness of it. Whatever you think AI is for, that was a version of the answer.
I keep coming back to that moment. A person who could not move was making something. The substrate underneath was an EEG signal and a model and a robot, and the thing on top was someone painting.
A thousand-person conference that didn’t exist a few months earlier, ending in a room of people quietly watching a woman paint with her mind. That’s the version of agency I want to remember on both sides.