Full Stack & AI Engineer building production systems that actually work.

AI Regulatory Document Processing Platform

I built a pipeline that uses LLMs and rule-based parsing to pull data from messy PDFs. It turned a task that took weeks of manual review into something that takes hours.

Challenge

Compliance teams were manually digging through dense documents for chemical data, which was slow and prone to errors.

Solution

I automated the extraction using document chunking and strict prompt validation, with a review queue for human supervision.

Outcome

Cut manual review time by 80% and expanded the knowledge base by 50%. It's now a core part of the product.

Automated Localization Pipeline

I built a system to extract and translate 1,200+ scattered strings from a large codebase. It uses a custom regex extractor and a DeepL pipeline to handle the bulk of the work.

Challenge

A cancer-screening platform was stuck because over 1,000 hardcoded strings were blocking it from launching in new markets.

Solution

I evaluated LLMs for extraction but found a custom regex solution was more accurate for code. I then built a DeepL pipeline to automate the translations.

Outcome

Saved 100+ engineering hours and roughly €7,500 in localization costs. Enabled rollout into Latvian.

Patch-Based Retina Recognition

A prototype for identifying people using small patches of a retinal image. I used a convolutional autoencoder to learn unique patterns without needing perfectly labeled data.

Challenge

Retinal vasculature is unique but high-variance, making it hard to identify individuals across different images.

Solution

I trained an autoencoder to learn patch embeddings and built a scoring function to find matches based on those features.

Outcome

Final grade of 8/10. It’s a project I still revisit when thinking about representation learning for medical images.

ZZP Risico (zzprisico.nl)

Hypex