How We Built a Privacy-First DPP Platform with Mistral 7B

The Technical Challenge

Generating a Digital Product Passport (DPP) from unstructured PDFs is incredibly hard. Regular Expression (Regex) scraping fails because every manufacturer formats their spec sheets differently. We needed an LLM (Large Language Model) that could intelligently read and map unstandardized text to strict EU JSON schemas.

Why Local AI?

We briefly prototyped with OpenAI APIs, but immediately realized the dealbreaker: B2B clients will not upload pre-release product blueprints to a public cloud API. We had to go local.

Our Architecture Stack

1. Inference Engine: We use Ollama running the mistral:latest (Mistral 7B) model. It strikes the perfect balance between high reasoning capability (to follow strictly defined JSON schemas) and low VRAM/CPU footprint, allowing edge-server deployments.

2. Asynchronous Job Processing: Extraction takes 10 to 60 seconds depending on the PDF complexity. We implemented a FastAPI backend that queues the extraction task to a Celery Worker via a Redis message broker. This ensures our API is never blocked and users get real-time polling updates.

3. Frontend Experience: The user interface is crafted with Next.js 14 (App Router) and styled using Tailwind CSS's darkest, most premium color palettes. React-Markdown provides smooth rendering of the final passport.

By combining Celery, Mistral, and Next.js, we managed to build the most secure, privacy-first DPP platform on the market.