Challenge
The client provides market intelligence on construction projects. Its basis is a huge volume of texts, documents and blueprints. The faster and more accurately they are processed, the more current and valuable the product.
Solution
We developed and launched a processing pipeline with eight custom deep learning models:
- Classification of incoming documents.
- Information extraction and deduplication.
- Generative AI enrichment and churn analysis.
Technically, the pipeline is built on Python, PyTorch, Hugging Face and NLP methods, complemented by generative models.
Result
- More than 300,000 documents per month run through the pipeline.
- Around 90% average model accuracy with minimal latency.
- Faster, less error-prone processes - and more current market intelligence for the client’s customers.
* Project: we owned the delivery · ● in operation: we run the solution today
