World Bank — Romania.
A two-part data engagement for the World Bank in Romania's energy sector — turning ~100,000 building energy certificates (dozens of gigabytes of photos, scans and spreadsheets) into one structured database, and turning the national census into accurate maps of how the country heats its homes.
- Industry
- Public sector / document intelligence + geospatial
- Timeline
- 2025
- Outcome
- ~100,000 documents → one database · 42-county map set
The problem
The World Bank's work on Romania's energy and heating sector hit the same wall twice, at two very different scales. At the building level there were roughly 100,000 energy performance certificates — plus supporting spreadsheets, photos and scans, dozens of gigabytes in all — in every conceivable shape, format and version. Many weren't even data: they were photographs of documents and OCR-quality PDFs. At the national level, the official 2021 census was authoritative but unmappable: its most detailed file reported over 38 million people for a country of 19 million — a near 2× overcount caused by the way administrative units nest inside one another. In both cases the information existed; it just couldn't be queried, compared or trusted.
What I built
Two pipelines that turned unusable source material into clean, decision-ready data.
A national energy-certificate database. A document-intelligence pipeline that reads roughly 100,000 certificates across their many formats and versions — using LLMs extensively to pull structured fields out of photos and OCR'd PDFs, not just clean files — then normalises and consolidates everything into a single Postgres database, with CSV exports for analysts. Dozens of gigabytes of scattered documents became one queryable source of truth.
An accurate national heating picture. On the census side, I traced the 2× overcount to its root cause in the administrative coding and built a filtering method that brings the national total back to within 2% of the official 19.05 million — with all 42 counties intact. From there: extraction from dense, multi-level Romanian-language tables, translation into clear heating categories, matching to county boundaries at 100% coverage, and a set of 11 print-ready maps — district heating, natural-gas dependence, wood and biomass, energy poverty, the dominant technology per county, and the urban–rural insulation gap — with full methodology and quality-assurance documentation.
Both were built in the agentic model, with Claude Code.
The outcome
From two piles of unusable data — a hundred thousand mixed-format documents and a census that didn't add up — to a structured database and a map set a government can act on, the latter delivered ahead of its November 2025 deadline. The maps made the policy picture concrete: roughly 493,000 homes (5%) have no heating system at all, concentrated in the rural south and east, and a quarter of all dwellings have no thermal insulation, with rural areas more than three times worse than cities. Together they gave the World Bank a factual foundation for discussions on Romania's energy sector — at both the building and the national level.