Turning long renewable energy proposals into clean, structured data with AI that runs privately.

The challenge

Renewable energy proposals are long, inconsistent documents. The facts that matter, such as who is behind a project, where it is, how big it is and what stage it has reached, are scattered across dozens of pages in different formats. Reading them by hand is slow, and copying details into a spreadsheet invites mistakes.

The documents can also be commercially sensitive, so sending them to an outside AI service was not an option. Whatever we built had to run privately.

Our approach

• Ran open-weight language models locally through Ollama, so documents never leave the environment they are processed in.
• Split extraction into several focused passes instead of one giant prompt. Each pass looks for a small group of related fields, which proved far more reliable on long documents.
• Defined a strict JSON schema for the output, so every proposal produces the same shape of data, ready for a database or spreadsheet.
• Iterated through more than twenty versions of the pipeline, testing each against real proposals and tightening the prompts and validation as we went.

What we learned

Most fields came out cleanly once the work was split into passes. Organisation names were the stubborn exception: the same company can appear under a trading name, a parent company and a project vehicle in one document. That is exactly the kind of field where a quick human check adds the most value, so the pipeline is designed to flag uncertain values for review instead of guessing.

The goal is not to remove people from the process. It is to give them a clean first draft in seconds.

Where this applies

The same pattern works for any business buried in documents: supplier invoices, contracts, planning applications, application forms or compliance paperwork. AI reads the document, fills in the fields, and people only look at the parts that need judgement.