Diploma Thesis Rant about AI

I recently dug up my diploma thesis from HTL Weiz again (“Untersuchung KI und praktische KI Anwendung”, 2023/24) and thought it deserves a proper post. Anyone who wants all 34 pages of math / screenshots: the PDF is here.

Why AI of all things

I’m an (Open-Source) AI enthusiast, plain and simple — ever since I stumbled over the text-davinci models in 2017, long before the ChatGPT hype wave. So when the diploma thesis came up the topic was honestly the only option.

The project partner was Learn4D365, an online learning platform for Microsoft Dynamics 365 with over 340 courses. Their existing “chatbot” was a little navigation popup with exactly zero AI in it. 340+ courses and no smart way to search through them — basically the perfect playground.

So I built SAINDY — Self Service AI for Learn4D365. A support bot that actually reads the course data and helps you find what you need.

The stack

PHP, vanilla JavaScript and the OpenAI API. No fancy framework and no (real) vector database — the “database” was a JSON file. The client wanted the simplest thing that runs on a basic PHP host. Essentially a homemade way to let the AI execute functions — back then features like that were only just getting started.

How it works — basically:

• The user asks a question and the backend embeds it with an OpenAI embedding model → you get a vector
• That vector gets compared against every stored course vector via cosine similarity
• The top 5 matches come back and on request the bot answers the question with the found courses as context

What does it actually cost

With gpt-4-0125-preview back then: worst case for a full interaction — search plus answer with context — around $0.043. In practice while testing: under a cent. And the API prices dropped to about a quarter within a year.

Looking back

Would I build it the same way today? Definitely not — I’d reach for a real vector DB, add streaming responses and probably not use PHP. But as a proof of concept it did exactly what it was supposed to: all requested features built and shipped, hosted as a beta on saindy.io — and one very happy client.

Got a pile of data and a vague feeling that AI could help? That’s exactly where Learn4D365 started too: