Acquisition pipeline

Blue Elephant Energy GmbH (solar PV asset manager)

Millions

Websites the crawlers search for relevant notices

3 months

Solo engagement in 2024, from architecture to a pipeline in production

Live

The pipeline has run independently at the client since the project closed

Challenge

Too many notices

Tenders and official announcements are scattered across the web, in every imaginable format and some only as scanned PDFs. Working manually, the acquisition team cannot keep up with the volume, even though it is precisely the details that decide how a park is valued.

Evaluate, don't collect

An end-to-end pipeline that finds relevant notices, makes them machine-readable and prepares them so the team can evaluate efficiently.

From noise to signal

Millions of websites parallel web crawlers
Notices as text OCR with AWS Textract
False positives removed rule-based pre-sorting
Relevant information structured LLM analysis
Team evaluation acquisition tool

The inexpensive stages filter first, so the LLM only sees what is worth processing.

Solution

Web crawler architecture

Parallel, high-performance API extraction processes search millions of websites for relevant publications.

OCR stage with AWS Textract

Scanned and PDF-based notices become machine-readable text, the foundation for the LLM stage.

Algorithmic pre-sorting

Rule-based filters eliminate obvious false positives before the language model runs.

LLM analysis

A language model extracts the information relevant to the acquisition and structures it for evaluation.

Acquisition tool

A frontend in React, Next.js and Tailwind where the team works with the prepared data.

Result

Running independently in production.

The acquisition team processes notices at a scale that would be impossible manually. The pipeline was cleanly handed over to the client and has kept running without alm since the project closed.