Personal Projects
Private Finance Dashboard
A local, privacy-first budgeting app that reads bank statements (CSV, Excel, PDF) straight from your computer. No cloud, no bank login, nothing ever uploaded. It parses statements from four different banks across two countries, categorizes spending, and tracks savings goals.
The part I'm most proud of: a monthly spending forecast built on least-squares linear regression per category, trained only on completed months so the current partial month doesn't bias the trend. It falls back to a plain average when the fit is too weak (r² < 0.3) or history is too short. A small but deliberate choice, to avoid presenting a shaky trend as a confident number.
arXiv Research Agent
An autonomous agent that monitors new arXiv papers in machine learning and NLP, summarizes them, and scores their relevance against a set of research interests: search, fetch, extract, summarize, and score, using real tool-calling rather than a single prompt-and-respond wrapper.
Academic Projects
Motivated Reasoning in LLMs
An empirical study of whether value-conditioned language models fairly update their positions when presented with verified contradictory evidence, or whether they acknowledge that evidence and then dismiss it through seemingly reasonable rhetoric. The project combines verified argument datasets, human-coded debate examples, controlled persona experiments, and qualitative evaluation of evidence-response strategies.
An initial 90-case study revealed that numerical confidence can shift while the underlying reasoning remains effectively unchanged.
PikoGPT: A Compact GPT Built From Scratch
A 37M-parameter transformer built from scratch with a team of four: architecture, pre-training, and fine-tuning, with no pre-trained weights. I worked on the post-training stage, comparing two supervised fine-tuning strategies.
Hands-on experience in deep learning, training optimization, and model benchmarking.