Fine Print: an AI analyst on every listing

Buying an investment property means either relying on the selling agent or stitching together five different sites. Most people go with gut feel, and more than half of Australia’s landlords made a rental loss last year.

The hackathon brief: put useful intelligence where buyers already are, on the listing itself, using Microburbs data.

I formed and led a team of three, set the product direction, and built the agent system. Press F on any REA or Domain listing and a panel slides in, scoring the property against the buyer’s own brief: budget, deposit, minimum yield and hold period.

Under the hood, I ported our agent runtime to run inside the browser: the transcript is the only state, the runtime owns the loop (assemble, generate, call tools, repeat), and every Microburbs endpoint is a declared tool with an allow-listed host and a stored secret, so the agent can only reach what it is meant to.

The model never does the maths. Yield, weekly cost and supply timing are calculated in code, and every figure carries a tooltip showing where it came from. The analyst answers questions with a short reply and a visual card, and remembers every property you’ve looked at so it can compare them.

A working extension in a day: likely sale price from the agency’s quoting record, our own rent model (a $42 a week typical miss, about half the benchmark’s $86, across 65,777 City of Sydney apartments), weekly holding cost, new supply timed against your hold, hazards, a shortlist spreadsheet, and an analyst you can question.