At Accenture, I developed and validated Python components for an invoice-extraction pipeline using PyMuPDF, PaddleOCR, and local LLMs and VLMs. I also worked on deterministic validation and a containerized FastAPI service.
I benchmarked seven quantized models with llama.cpp and Ollama, comparing extraction quality, latency, and memory use. The best configuration reached 95.3% accuracy across 171 values from nine invoices.
For my MSc thesis at POLITECNICO DI TORINO – POLITO, I developed a LangGraph diagnostic agent for document visual question answering. I evaluated prompting strategies and vision-language models on corrupted questions derived from DUDE, without fine-tuning.

