RAG chatbots
Assistants for the University of Rijeka and the R.S. Bonum accounting firm. They retrieve, rank and answer, and the Bonum one talks back in Croatian. Here a query finds its six nearest chunks.
FAISSLangChainOpenAITTS
// AI research · simulations · small machines
Hi, I'm Bojan.
I build things that think a little.
AI researcher at the University of Rijeka (FIDIT), working on spotting new topics in the news with LLMs. Below: RAG chatbots, an AI agent, neural networks from scratch, cellular automata and one very polite robot, most of them running live in your browser, so poke them.
The Lab
Live experiments. Every card is a tiny running version of the real project.
Lumen
My master's thesis: a self-hosted AI agent on local, open-source models. It searches the web privately, runs Python, sends email, draws images and answers out loud, all on one GPU at home. Point at a tool to ask for it next.
LangChainlocal LLMRabbitMQReactThinking…
Tin Bot
An Arduino robot with an LCD face, ported to a 3D web and Android app. Same eye bitmaps, idle animations and buzzer tunes as the real firmware.
ArduinoExpothree.js
Neural network from scratch
A small multilayer perceptron written by hand to really understand forward passes and backpropagation. Point at a node and more of the signal flows into it.
PythonNumPybackprop
Ishihara number recognizer
A machine learning model that reads the numbers hidden in colour blindness test plates.
computer visionML
Elementary cellular automaton
A purely functional Python implementation. Every time you point at the card it picks a new rule. Rule 30 makes chaos from a single cell.
functional Pythonrule 30
Interactive fractal tree
Recursion you can bend. Move across the card to change the branch angle. The original is built in Unity.
recursionUnity
Visual cryptography
Each share alone is pure noise. Stack them and the secret appears. Move across the card to slide them, they meet in the middle.
Pythoncryptography
Image to ASCII converter
Turns any image into text art, with adjustable size and brightness. Here it's the classic spinning donut: move across it and the lens reveals another famous donut.
PythonPillow
Home web server
An old PC that runs all of this. Apache sits in front: this site and Tin Bot are plain files, the apps run behind it, and the tools that need a server get their work from a small API through RabbitMQ. Each dot is a request, and its answer comes back in orange. Point at a service to send it traffic.
ApacheDockerRabbitMQLinux0 requests served
Apps for real people
Things built for people to actually use, not just to look at.
Archive
Older projects, from Delphi to Unity.
About me
I'm from Ližnjan in Istria and live in Rijeka, where I'm a researcher at the Semantic Technologies Laboratory of the Faculty of Informatics and Digital Technologies (FIDIT), the faculty where I finished both my degrees cum laude. On the MEDIANA project I work on an AI system that follows the media, spotting new names and topics as they appear with LLMs, embeddings and knowledge graphs. Before that I was at Ericsson Nikola Tesla, first as an intern and then as a full-stack developer, and mentored two of its summer camp teams.
I've been programming since I was a kid, starting with games in Delphi, and these days I mostly build AI systems. This site, the chatbots and the other apps all run on a home server I set up myself.
Languages
Python, TypeScript, JavaScript, C++
AI & data
LLMs (local and OpenAI), LangChain, RAG, Hugging Face Transformers, PyTorch, FAISS, zero- and few-shot learning, NLP
Frameworks
FastAPI, Django, Flask, React, React Native, Vue.js
Infrastructure
Docker, RabbitMQ, PostgreSQL, MLflow, Apache, Linux
Research notes
How the things I research and build actually work, explained with pictures.
How RAG worksHow a chatbot answers from your own documents instead of from memory, step by step, with the settings behind Emma.AI agents: a language model with tools and a loopWhat turns a chatbot into an agent, and the ReAct pattern that lets it decide which tool to use, from my master's thesis.Spotting emerging topics with BERTrendHow a pipeline of embeddings and clustering notices a new story while it's still small, and tells it apart from noise and from topics everyone already covers.
All 9 notes →