Parseur

The Challenge

Extracting structured data from unstructured documents (PDFs, images) is notoriously hard. Traditional OCR pipelines are brittle and require manual rules per document type. I wanted a system that could understand any document layout using LLMs, with auto-correction and validation.

The Solution

• Multi-LLM pipeline combining OpenAI GPT-4o and Anthropic Claude 3.5 Sonnet
• Intelligent extraction with auto-correction and Zod schema validation
• S3-compatible storage with MinIO for local development
• Background job processing with Inngest for async document pipelines
• Full-stack Next.js 16 app with Prisma ORM and PostgreSQL

Results

Learnings

Multi-LLM orchestration taught me that no single model excels at everything. GPT-4o handles visual layout understanding better, Claude excels at structured reasoning. The key is composing them — not choosing between them.