I am a scientist that builds systems that are traceable, evaluated, and built to be trusted.
AI/Data engineer · Arnhem · Open to opportunities
Scroll ↓A full retrieval-augmented generation system for scientific papers, built from scratch with an integrated frontend and backend on open-source Python. The goal from the start was speed and validation: answers should come back fast, and every answer should be checkable against the exact source sentence in the original PDF, so the system is built to be fully transparent, not just trusted. Below is a demonstration of its core features.
Users upload papers into a shared library, and speed was a deliberate focus from the start: a document is fully processed and ready to query in under 30 seconds. That speed is what makes the tool frictionless to use, there's no waiting around, no reason to hesitate before uploading the next paper.
Ask a question and get an answer grounded in the source material. Every claim links back to the precise sentence it came from, not just the page or paragraph. You can verify any answer in a second instead of taking it on faith.
Not every paper follows a clean, predictable outline. A custom-trained machine learning model handles that reality directly, detecting sections even in documents with irregular or inconsistent structure. Rather than failing outright when a paper breaks the usual pattern, the model degrades gracefully, still surfacing the most likely section boundaries so an answer can always be traced back to where it came from. If the detected section isn't right, the user can validate it immediately: selecting a source scrolls straight to that section in the paper, and if the detection missed, the user can simply select the page instead.
Nothing about this system is tied to scientific papers specifically. The same pipeline, library, retrieval, and traceability layers can be pointed at any domain with documents worth trusting: legal contracts, medical records, internal policy, financial filings. The architecture doesn't change; only what gets indexed does.
It's also built entirely on open-source technologies, with no dependency on any proprietary platform. That means it can run completely on European infrastructure, outside the reach of the US Cloud Act, if that's what a given deployment requires. It was architected with this in mind from the start: depending on how much trust and control a domain demands, it can scale from a standard cloud deployment all the way to a fully sovereign, self-hosted solution, without changing the system itself.