Australian Job Market Visualiser

An Australian adaptation of Andrej Karpathy's US Job Market Visualizer, rebuilt with real Australian labour market data from government sources.

View the live visualisation

What was adapted

The original repo scraped US Bureau of Labor Statistics (BLS) occupation data, scored AI exposure via an LLM, and rendered an interactive treemap. This fork replaces the entire data pipeline with Australian equivalents:

Original (US)	Adapted (Australia)
Bureau of Labor Statistics (BLS)	Jobs and Skills Australia (JSA)
SOC occupation codes	ANZSCO 4-digit unit groups (361 occupations)
USD pay figures	AUD median weekly full-time earnings x 52
US degree classifications	Australian Qualifications Framework (AQF)
BLS 10-year employment projections	JSA/Victoria University 5-year projections
143M total employment	14.4M total employment
24% of jobs in declining occupations	5% of jobs in declining occupations

What's here

An interactive treemap where each rectangle's area is proportional to total employment and colour shows the selected metric — toggle between projected growth outlook, median pay, education requirements, and AI exposure. Click any tile to view its full JSA profile.

Data sources

All data is from official Australian government publications:

• JSA Occupation Profiles (Nov 2025) — Employment levels, median weekly earnings, demographics, education attainment for 358 ANZSCO 4-digit occupations
• JSA Employment Projections (May 2025–2035) — 5 and 10 year growth projections by occupation (Victoria University modelling)
• ABS Employee Earnings and Hours (May 2025) — Average weekly earnings by 4-digit occupation (DO011), used as fallback for 36 occupations missing JSA earnings

The raw Excel files are in data/.

Data pipeline

1. Build occupation list (build_occupations.py) — Scrapes the JSA occupations index to build occupations.json with ANZSCO codes, titles, categories, and URLs.
2. Scrape (scrape.py) — Playwright downloads raw HTML for all occupation profile pages into html/.
3. Parse (parse_detail.py, process.py) — BeautifulSoup converts raw HTML into clean Markdown files in pages/.
4. Tabulate (make_csv.py) — Extracts structured fields into occupations.csv.
5. Score (score.py) — Sends each occupation's Markdown description to an LLM with a scoring rubric (0–10 AI exposure). Results saved to scores.json.
6. Build site data (build_real_data.py) — Parses the JSA and ABS Excel files directly and merges with AI exposure scores into site/data.json.
7. Website (index.html) — Interactive treemap with four colour layers: Growth Outlook, Median Pay, Education, and Digital AI Exposure.

LLM-powered colouring

The "Digital AI Exposure" layer estimates how much current AI will reshape each occupation. But you could write a different prompt for any question — e.g. exposure to humanoid robotics, offshoring risk, climate impact — and re-run the pipeline. See score.py for the prompt and scoring pipeline.

Caveat: These are rough LLM estimates, not rigorous predictions. A high score does not predict a job will disappear. Software developers score 9/10 because AI is transforming their work — but demand could easily grow as productivity increases. The scores do not account for demand elasticity, latent demand, regulatory barriers, or social preferences for human workers.

Key files

File	Description
build_real_data.py	Parses government Excel files into site/data.json
occupations.json	Master list of 361 ANZSCO 4-digit occupations
scores.json	AI exposure scores (0–10) with rationales
data/	Raw government Excel data files (JSA, ABS)
site/	Static website (treemap visualisation)
index.html	Treemap (also at repo root for GitHub Pages)

Setup

uv sync
uv run playwright install chromium

Requires an OpenRouter API key in .env for AI exposure scoring:

OPENROUTER_API_KEY=your_key_here

Usage

# Build site data from government Excel files (no API needed)
python build_real_data.py

# Serve the site locally
cd site && python -m http.server 8000

# Or re-run the full pipeline from scratch:
uv run python build_occupations.py   # build occupation list
uv run python scrape.py              # scrape JSA pages
uv run python process.py             # parse HTML to Markdown
uv run python make_csv.py            # generate CSV
uv run python score.py               # score AI exposure (needs API key)
uv run python build_site_data.py     # merge into site/data.json

Methodology differences vs US version

The Australian visualisation looks noticeably more optimistic than the US version. Key reasons:

• Projection horizon: Australia uses 5-year growth projections (May 2025–May 2030) while the US uses 10-year BLS projections. The magnitudes are roughly comparable (+6.6% AU 5yr vs +3.4% US 10yr), but they measure different time spans.
• Fewer declining occupations: Only 5% of Australian jobs are in occupations with negative 5-year growth, vs 24% for US 10-year projections. This reflects Australia's strong population growth from immigration, which sustains demand across most occupations. Even over 10 years, only 13 of 358 Australian occupations are projected to decline.
• The projections do not account for AI: The JSA/Victoria University model projects employment based on historical trends, demographics, and industry structure. It does not currently reflect the potential labour market impact of generative AI adoption.
• Currency: AUD pay figures are not directly comparable to USD. At current exchange rates (~0.65 USD/AUD), an Australian median of A$92K is roughly US$60K.
• Education system: Australia uses the Australian Qualifications Framework (AQF) — Certificate I–IV, Diploma, Bachelor degree, etc. — rather than US degree classifications (Associate's, Bachelor's, Master's). The education groups are not directly comparable.
• Coverage: 332 of 361 occupations have real earnings data (296 from JSA, 36 from ABS Employee Earnings). The remaining 29 (mostly farmers, defence, niche trades) are estimated from ANZSCO skill levels. All 361 have AI exposure scores — 332 via LLM, 29 scored manually using the same rubric.

Notes on Australian data

• Currency: All pay figures are in Australian dollars (AUD). Median weekly full-time earnings are converted to annual by multiplying by 52.
• Classification: ANZSCO (Australian and New Zealand Standard Classification of Occupations) 4-digit unit groups. 8 major groups: Managers, Professionals, Technicians & Trades, Community & Personal Service, Clerical & Admin, Sales, Machinery Operators & Drivers, Labourers.
• Employment projections: JSA / Victoria University Employment Forecasting model (May 2025–May 2035). These do not currently reflect the labour market implications of generative AI adoption.
• Education: Uses the Australian Qualifications Framework (AQF) — from Year 10 through Doctoral degree — rather than US degree classifications.