→ All work

Live · Oct 2026

Winner · Hong Kong AI Summit hackathon 2026

GridReady

Drop a data centre on Hong Kong's power grid and watch a whole day play out. The map shows what overloads and when, what that costs, and the cheapest fix. Built in 36 hours for CLP's track at the Hong Kong AI Summit hackathon, and it won.

Open source · View on GitLab
Role
Led the build, orchestrating an AI-agent team
Stack
TypeScriptReactMapLibreWeb WorkersDC power flowOpenStreetMapClaude API
Numbers
90 nodes · 151 lines48 power flows a day36 hours, idea to live
Verify
Live site ↗Source code ↗
The 90-second demo we pitched with · sound on

GridReady · the story

Land secured. Can the grid supply it?

Scroll

Winner
HK AI Hackathon 2026, with a HK$7,500 prize from HKUST MBA Programs
600
people registered, across Finance, Urbanization and Startup tracks
90 s
video submission first, then a 3-minute live pitch battle
36 h
from a blank repo to a live URL the judges could scan

A day on the grid, twice.

Flat AI-training load · Hong Kong summer evening peak · day two with the fix

A data centre draws almost the same power at 3am as at 7pm, so the line feeding it is over its rating all day. Four new transformers bring the worst hour from 258% down to about 90%.

Try it.

Live · drag a data centre onto the map, then Review and Fix

Live · the same build the judges used Source on GitLab ↗ Open full screen ↗

Start with the one-click example, a 300 MW AI campus at Sandy Ridge. Zoom into a district to see local 11kV substations and spare capacity.

From the pitch.

Clips from the demo video

Only 107 MW fits. The substation hits 258% of its rating, about HK$50M a day at risk
Fixes ranked cheapest first. Apply them and the whole day is simulated again
Day 2, HKUST Business School Central: the winner's cheque
Day 2, HKUST Business School Central: the winner's cheque

Recorded from the local build the night before the pitch

How an answer is computed

PLACE

a data centre at a substation, with its MW and load shape (flat, AI training, cloud)

FLOW

DC power flow over the CLP network for 48 half-hour steps, in a Web Worker

BREAK

every line and transformer over its rating, and when

COST

transformer wear for CLP, outage risk for customers, in HK$ per day

FIX

upgrades, transfers, phasing or a different site, ranked by cost and lead time

Rules first, so the demo cannot invent a number. Claude only rewrites the brief, and every figure is checked.

The worst hour at Sandy Ridge

As is

258% of rating · about HK$50M a day in wear and outage risk

With the fix

90% · four transformers, about HK$180M and 30 months

Indicative, from public data and estimated ratings

GridReady · the write-up

How it was built.

A hackathon of 600

The Hong Kong AI Summit hackathon, hosted by HKUST Business School, with about 600 people registered. We took CLP’s Urbanization track. Two rounds: a 90-second video, then a 3-minute pitch battle. GridReady won, and with it the HK$7,500 prize.

An answer in seconds

Data centres are Hong Kong’s fastest-growing load, and grid connection is the bottleneck. A developer can buy land and wait months to learn whether CLP can supply it. GridReady answers in seconds, for both sides.

A real grid: 90 nodes

The 400kV backbone and substation sites are real, from OpenStreetMap: 90 nodes, 151 lines, 7,477 MW at peak. The 132kV layer and the ratings are estimates, and the app says so on every screen. A DC power flow solves every half hour of a summer weekday in a Web Worker.

AI that cannot invent a number

Verdicts, problems and fixes are rule-based. Claude only rewrites the brief into plain English, and if any figure in its answer is not in the computed facts, the rule-based text is shown instead.

What an overload costs

Costs come from IEEE C57.91 transformer ageing and the value of lost load, which turns a red line into a figure a business acts on.

Built with AI agents

I ran it as an orchestrator of AI agents: one wrote the engine against a single contract file, others built the data and UI on separate branches. GitLab CI deploys main to this site.

End of GridReady

Next project

1D Schrödinger scattering

An interactive quantum scattering simulator in the browser. Pick an energy and a potential, and watch the wavefunction, |ψ|², transmission and reflection update as you drag.

Next page