Project

Description

This school project, completed during the Master in Applied Artificial Intelligence, was developed for the Municipality of Amsterdam to help government officials simplify their letters for citizens with A1 to B2 language proficiency.

The application used Optical Character Recognition (OCR) to extract text from government letters, followed by Natural Language Processing (NLP) techniques to simplify the content. The pipeline included preprocessing, text simplification using Dutch pre-trained language models such as Geitje 7B Ultra and Llama 3.2 3B Instruct, and guardrails to verify model outputs.

For training and evaluation, metrics like BLEU, BERTScore, and SARI were applied, with further fine-tuning performed using Low-Rank Adaption (LoRA). The final application featured a React.js frontend and a Flask backend.

Through this project, I learned about the challenges of designing a user-friendly interface for citizens with lower literacy, the complexity of defining what counts as a “difficult” word, and the difficulty of improving already highly optimized Dutch language models.