How I built a China gas dashboard, and why it's a low-hanging fruit for AI agents

In my last post I asked whether an AI agent can take on a real commodity workflow — not just summarize news or scrape a website, but work through messy external information, update a model, and explain its reasoning.

The dashboard I built this summer is a good example of the middle ground: a real analyst deliverable that is also an easy target for an AI agent.

The dashboard

The live China gas explorer now sits inside this website. It opens as a guided explanation and can switch into a compact dashboard for repeat visits.

It shows China’s monthly natural gas actuals from the JODI database, arranged by gas year (October to September), with rolling five-gas-year comparisons, an import stack for LNG versus pipeline, a sector-demand model anchored to Carbon Monitor China data, and a deterministic storage optimizer on the TTF curve.

Why it’s a low-hanging fruit

The whole thing is static. No server, no database, no API keys needed at runtime. The heavy lifting is:

Public data: JODI gas actuals, Carbon Monitor China, AGSI storage — all freely downloadable.
Build scripts: a Node script downloads the JODI zip, extracts China rows, and writes a data.js file. Another script turns Carbon Monitor data into the sector model.
Plain HTML/CSS/JS: the pages read those data files and render charts in the browser.

That shape — fetch public data, transform it, render it statically — is exactly what an AI agent is good at. There is no proprietary input, no live authentication, and no real-time market feed to defend. The main judgment is deciding what to show and whether the numbers reconcile, not how to fetch them.

What it tested

Building it from scratch with an agent meant the model had to:

Find and understand a government database schema (JODI flow codes).
Keep units and conventions straight (bcm, gas years, LNG vs pipeline).
Decide how to structure the data so the charts would be honest — e.g. not double-counting, checking the residual balance.
Produce something a non-programmer can actually read.

That’s more than a morning news summary, and less than a full trading view. It’s the kind of task where the analyst still sets the direction and sanity-checks the output, but the agent does the hours of fetching and formatting.

The takeaway

A lot of commodity analysis is still done as if static dashboards were hard. They’re not. When the data is public and the output is a static page, an AI agent can assemble the whole thing quickly — leaving the analyst with the part that matters: deciding whether the picture is right, and what it means.

If you’re curious whether AI can do real analyst work, building a dashboard from public data is the cheapest possible experiment. And it’s one most people can just try.

The finished China gas explorer is available as a permanent, date-free page.