On 10 October 2026, I took ResearchAgent to the main stage at AI Tinkerers Singapore. I presented a live demo at Singtel 8George: a research agent with a wallet, an evidence gap to fill, and a fixed budget. The project began with our SingHacks team, Nguyễn Trần Thành Minh, Jiamin Wang, Andrew Nguyen, and me. We had reached the finals there in September. This was the next step: bringing the evolving project in front of a room of builders.

Give the purchase a reason

The walkthrough starts with free sources. The agent builds an answer, identifies what it still cannot support, and considers whether a paid article could improve it. The question is how much useful evidence that article adds. Another version of something already known should not earn a purchase just because its title looks relevant.

On the projected screen, you can follow the question, the available budget, the chosen source, and the answer after the purchase. Those details make the decision inspectable. The model can propose a source, while budget limits and payment checks constrain the action. Delivery verification then checks what arrived before the result becomes part of the answer.

A presenter explaining the projected ResearchAgent diagram, with separate reader, writer, and testnet ledger layers.
Walking through the reader, writer, and payment layers. Photo: event attendee, via AI Tinkerers Singapore’s 10 October gallery.

The demo uses fictional writers and articles, with XRPL Testnet settlement. No real money moves. That boundary matters when watching the presentation: the prototype lets us examine spending controls, delivery proofs, and recovery from failed purchases in a contained setting. A verified delivery still does not establish that an article’s claims are true. It establishes what was delivered against the promised evidence.

A Cloud Run demo worth taking home

The live demo ran on Google Cloud Run and placed among the top three Cloud Run demos at the meetup, earning US$100 in Google Cloud credits. I shared that result in my post after the event. It was a separate milestone from the SingHacks finalist result, and presenting the work itself was a meaningful part of it.

I also turned the build into a retrospective: 39 days, 248 commits, 75 pull requests, and 422 prompts to two coding agents. The charts are a record of how this project was made. My favourite slice was one night with 23 pull requests between 1 am and 3:19 am. The writeup, slides, and retained walkthrough let someone explore both the engineering and the path to the presentation.

Follow the trail