Question
Does structure and chronology move relevant evidence closer to the top of the results?
Mateusz Wiecek / My story
Curiosity connects the chapters.
Sport, people and AI. Different settings, with the same wish to understand how things work and get a little better at them.
University of Strathclyde
I have completed the requirements for the MSc Artificial Intelligence and Applications at the University of Strathclyde. The formal award is pending.
Before the models, there was the pitch.
My physical-education background at AWF, football and freestyle are part of how I approach learning. A trick rarely works on the first attempt. You break it down, notice the detail that failed and try again. That habit feels familiar when I am debugging code or investigating a model’s mistakes.
Sport is also a reason to step away from the screen. I want this site to show that side of my life alongside the technical work.
Good work starts with the people doing it.
Since 2012, I have worked seasonally with Smile Camp, leading groups and sports activities while also supporting web content and digital communication. It is a mix of organisation, responsibility and finding a clear way to explain the next step.
Operational work with Amazon DSP and Flex adds another perspective: plans meet changing conditions. Being useful means staying calm, noticing exceptions and taking responsibility for what happens next.
From physical education to artificial intelligence.
The MSc Artificial Intelligence and Applications at the University of Strathclyde brought a different set of tools: Python, machine learning, information retrieval and quantitative evaluation. I worked on neural networks from first principles, music-data modelling and retrieval over synthetic records.
The common thread is the quality of the question. A high score is a starting point for investigation: what was measured, against which baseline, and where does the result stop being useful?
I want the evidence to be easy to inspect.
My portfolio brings the code, method and limitations together. FireBid Radar extends that interest into product work: helping people find relevant information and see the evidence behind it. Movement Lab explores the connection between AI and an everyday interest of mine: movement.
I am interested in conversations about applied AI, evaluation, useful products and learning. Friends are equally welcome here: the Lifestyle pages are a place to discover the person behind the projects.
MSc dissertation / public summary
Finding the right evidence, in the right context.
My research asks whether the section of a clinical record and the time it refers to can improve retrieval over a lexical baseline. All compared variants keep the same patient filter. The controlled study uses MIMIC-IV-Note v2.2 discharge summaries; these are not NHS records.
Method
Compare a BM25 baseline with section-aware, time-aware and combined reranking. Evaluate retrieval separately from answer generation.
What I learned
In the defined question set, the combined variant improved the rank of relevant evidence. The evidence for higher Recall@5 was less decisive. One metric cannot tell the whole story.
Research prototype with a small, single-annotator question set. This summary does not claim clinical effectiveness, NHS compliance or deployment readiness. Patient records, questions and restricted data are not published here. The separate public Clinical RAG project uses synthetic data.
Explore the public synthetic-data project →