CASE STUDY
Museum XTD
My bachelor's thesis at HEIG-VD, early 2023: a chatbot that answered questions about MuseumXTD using only MuseumXTD's own documents, in French. Retrieval-augmented generation, a few months before the AI boom.
THE CONTEXT
The Challenge
MuseumXTD is a resource platform for museums working through their digital transition. It had grown into a large body of French-language documentation, and the practical problem was that the answers were all in there somewhere and nobody could find them. My bachelor's thesis at HEIG-VD was about fixing that.
A search bar returns documents, and the person asking still has to read them and work out which paragraph applies. What the platform needed was something that could do that reading itself and come back with an answer, in French, without stepping outside the material it had been given. In early 2023 there was no standard way to build that.
An assistant that answered only from MuseumXTD's own documents, in French, and invented nothing. In early 2023, none of that was off-the-shelf.
THE PIPELINE
How it was built
AI Integration
RAG Architecture
Proof of Concept
01. Restructuring the source material
The platform was a static site. I moved the documentation into Nextra so the Markdown was clean and consistent enough to be indexed reliably. Retrieval quality depends almost entirely on this step, and it is the least visible part of the project.
02. Building the pipeline
Tested Mendable and Flowise, then built a pipeline that kept the model restricted to MuseumXTD's own documents. In early 2023 that meant assembling it out of parts that had not been designed to work together.
03. UX and deployment
A chat interface built into the platform rather than bolted onto it, working in French and holding to the existing brand.
RESULT
What it actually did
Answers you could check
It stayed inside MuseumXTD's own material and cited its source on every reply, so when it got something wrong you could see which document it had misread. For an institution, that traceability mattered more than raw accuracy did.
Custom RAG Pipeline
Retrieval-augmented generation, built in early 2023, before the tooling for it existed.
Nextra & Markdown
Structured for both kinds of reader. The documentation ended up in one format that served the human reader and the model equally well.
The prototype
It stayed a proof of concept. It answered domain questions in French out of the platform's own documents, which was enough to show the approach held up for an institution that could not afford a model inventing things.
Early 2023
In early 2023 there was no playbook for custom AI integrations. You picked the tools, wired them together yourself, and kept whatever held up under real questions. Getting a grounded French chatbot working out of parts that weren't built to fit was the point where this stopped being a thesis topic and started being the thing I wanted to do. Most of what I built then has been superseded since, but the instinct for what these systems can and can't be trusted to do came out of doing it early, before there was a standard way. That head start still shapes how I approach AI work now.
Pushing what was technically possible at the time didn't just get me through my Bachelor's thesis, it's what convinced me this was the path I wanted to build my career on. Looking back now, it's wild to see how far AI has come in such a short time. I feel lucky to have started experimenting with it early.