Projects

Live evaluation, with benchmarks generated fresh on every run

A framework that generates new synthetic test data for each evaluation run, and an ablation study of when that data is a realistic stand-in for a real benchmark.

Completed
#llm
#evaluation
#benchmark-contamination
#python
#ablation

Homelab as code

Turning a hand-managed home server into Ansible and Docker Compose, with service routes, DNS and documentation generated from one inventory.

Ongoing
#ansible
#docker
#infrastructure-as-code
#ci
#self-hosting

BreathIn v2

A rework of my bachelor-thesis air-quality application, starting with a cleaner structure for the data pipeline and a containerised database.

Planned
#python
#postgresql
#docker
#uv
#air-quality

Blokus engine built with LLM coding agents

A Blokus game engine with a CLI and a web GUI, built by a team of four with LLM agents, and a record of which prompting guidelines held up and which failed.

Completed
#software-engineering
#llm-agents
#python
#hexagonal-architecture
#testing

Agentic web question answering, with an ablation study

A multi-agent RAG system that answers factual questions from live web evidence, and what each of its eight components actually contributes.

Completed
#llm
#rag
#agents
#evaluation
#ablation

Wine recommendation on 21M ratings

Classical, hybrid and graph-based recommenders compared on the X-Wines dataset, including cold-start users and wines.

Completed
#recommender-systems
#pytorch-geometric
#lightgbm
#graph-neural-networks

Does synthetic data help classifiers on a tiny dataset?

A controlled comparison of nine classifiers trained on real, SMOTE-generated and deep-learning-generated obesity-risk data.

Completed
#data-mining
#scikit-learn
#xgboost
#synthetic-data

BreathIn, near-live air quality for Czech cities

A web application and data pipeline that collects air-pollution measurements daily and tells residents how their city compares with its own history.

Completed
#django
#postgresql
#postgis
#data-pipeline
#air-quality