Shine Li

Learnings from building AI workflow/task system

[Shine Li @fcpga_te]

🚀 Launch Alert 🚀

BoomAU is live on Fazier

Identify%20next%20boom%20suburbs%20in%20Australia%20with%20high%20accuracy

Please support us here 👇

https://t.co/L5b1sFjbb7 ⁠

[Shine Li @fcpga_te]

Full breakdown — why DOM behaves as a binary, the four filters, how to read DOM in your own research without a scoring engine:

→ https://t.co/s7nNv1SQND

We score 393 Australian suburbs fortnightly using these filters. Wishlist: https://t.co/nIsa7SScdg

[Shine Li @fcpga_te]

In our 78-suburb boom backtest, days on market wasn't a score — it was a gate.

Every suburb that genuinely boomed was sitting under 45 days on market during its boom window. Not "lower DOM than average." Under 45. Full stop.

[Shine Li @fcpga_te]

Pass all four, you're in the scoring pool. Miss any one, you're out. No clever weighting saves a suburb that fails a gate. The v2.3 formula hits 85.7% detection accuracy and 0% false positives across the 78-suburb backtest — and most of that precision comes from the gates doing the work before the ranking ever runs.

[Shine Li @fcpga_te]

@joozio Not really any maintenance. If I see a bug, I just tell it to fix it. If I need something new, I just ask it to build it. e.g. I needed a file upload and copy&paste into tmux session on mobile. I just say so and it builds it then restarts the server. that's about it

[Shine Li @fcpga_te]

Building a Dashboard With 77 Agent Tasks

We built a workflow visualization dashboard entirely through AI agent tasks — 77 tasks from kanban board to WebSocket live updates to initiative-aware navigation. The pattern that made it work: a discovery engine as the single source of truth. Zero hardcoded workspace assumptions. The engine reads .claude/ directory contents and derives everything — agents, tasks, rules, skills, relationships. When the workspace structure changed, nothing in the UI broke because nothing was hardcoded. Build dashboards around discovery, not configuration.

[Shine Li @fcpga_te]

Fortnightly scored list opens soon. Wishlist: https://t.co/nIsa7SScdg

[Shine Li @fcpga_te]

Full breakdown — the corridor maps, the data sources, how to check a suburb yourself without our formula: →https://t.co/pIQOGp0bSK

[Shine Li @fcpga_te]

That third line is the window. If you wait for the median to tell the story, you've already missed it. The buyers priced out of the expensive suburb next door have already shown up — you just don't see them unless you're watching inventory and time-on-market.

[Shine Li @fcpga_te]

When building https://t.co/nIsa7SScdg, Most property investors I talk to chase growth. A small number wait one suburb over.

[Shine Li @fcpga_te]

→ Days on market collapse in a suburb adjacent to one that just boomed.
→ Vacancy drops below 1.5%.
→ Median price barely moves. Yet.

[Shine Li @fcpga_te]

Most investors "know" this pattern. Almost none can point to where the front of the wave is today. The signature sits in data months before it sits in the price:

[Shine Li @fcpga_te]

That's the ripple effect in one sentence, and it's why Sydney's inner west booms rolled out to Marrickville, then Dulwich Hill, then Campsie. Why Melbourne moved from Fitzroy to Brunswick to Coburg. Why South-East Queensland is grinding north along the rail line right now.

[Shine Li @fcpga_te]

I wrote up the full analysis with the state-by-state breakdown and the traps to avoid: https://t.co/47wMDRYPfA

[Shine Li @fcpga_te]

We ran 12,360 postcode-months through a walk-forward backtest. Cancelled the market tide so rising-tide years couldn't inflate the results. Checked which features actually predicted outperformance suburb-by-suburb, not just across time.

[Shine Li @fcpga_te]

There are traps. Thin markets with 15 sales a year produce noise, not signal. Single-employer towns crash when the mine closes. Cheap and declining is just cheap. You need affordability AND demand signals AND liquidity. Two out of three isn't enough.