AI evaluation
Model comparisons, answer-quality tests and error analysis. The outcome: a clear view of what a system can and cannot do.
↗Mateusz WiecekAI · Lifestyle
AI evaluation · Machine learning · RAG
AI you can
put to the test.
I build and test AI systems, from document retrieval to image recognition. I show what works, where a system fails and what to improve, bringing experience from sport and leading teams to the way I work.
AI + LifestylePrecision. Energy. People.
Strathclyde, physical education and Smile Camp. Technical precision meets the energy of sport and experience working with people.
About
I build and evaluate AI systems that can be checked.
I have completed the requirements for the MSc Artificial Intelligence and Applications at the University of Strathclyde. The formal award is pending. My route runs from physical education, football and group leadership to building AI systems. Curiosity, patience and responsibility for the outcome connect those worlds.
- AI evaluation & QA
- Machine learning
- RAG & LLMs
- Computer vision
How I can help
From a question to evidence.
Looking for help with an AI project or someone to join your team? These are the problems I work on.
RAG and machine learning
Prototypes for document retrieval, classification and prediction. The outcome: a reproducible experiment with code and metrics.
↗Technology for people
I combine technical work with experience leading sports groups. I explain complex ideas clearly and turn them into practical next steps.
↗Evidence over promises
Results you can inspect.
Selected metrics from academic projects. Every number has a defined dataset, evaluation method and limitations.
How I work
Question. Experiment. Evidence.
I do not treat a model score as decoration. First I define what we are really measuring, then I build a repeatable experiment and present the result together with its limitations.
- 01A better question
- 02A reproducible test
- 03A clear conclusion
011.000Recall@3 · Clinical RAG
retrieval over a synthetic EHR dataset
Method and limitation +
How it was tested: Qrels and explicit P@3, R@3 and MRR for patient-specific retrieval.
Limit: This is not 100% answer quality or clinical effectiveness. The data is synthetic.
Project source ↗0245.84%Sign Language MNIST
primary result on original test labels
Method and limitation +
How it was tested: A feed-forward network built in NumPy, with regularisation, early stopping and error analysis.
Limit: Label alignment gives a diagnostic 88.62%; it does not replace the primary 45.84% result. This is an image benchmark, not camera-based gesture recognition.
Project source ↗036.52657Spotify popularity RMSE
reported regression-model error
Method and limitation +
How it was tested: CatBoost, baselines, feature engineering, validation, diagnostic figures and tests.
Limit: The result depends on the dataset and target scale; it is not a recommender-system score.
Project source ↗047.50%GA strategy · out-of-sample
with 9.99% maximum drawdown
Method and limitation +
How it was tested: Chronological train/test evaluation, transaction costs and several comparison baselines.
Limit: A historical experiment, not a forecast of future returns or investment advice.
Project source ↗Experience
Work, projects and experiences I contribute to.
From people and communication, through demanding operations, to reproducible AI systems — each experience reinforces responsibility for the outcome.
2012—present · seasonal
Smile Camp
Camp leader & web/digital supportSince 2012, I have contributed seasonally to the Smile Camp experience, combining camp and sports leadership with web content and digital communication support.
Meet the team↗2025–2026 · formal award pending
University of Strathclyde
MSc requirements completedArtificial Intelligence and Applications. From neural-network experiments to a dissertation on evidence retrieval in clinical records.
Explore my story and research↗2019—present
Amazon DSP / Amazon Flex
Delivery Associate / DriverHigh-tempo operational work requiring routing accuracy, procedural compliance, independent judgement and effective exception handling.
More than a calling card
Explore the work, direction and person.
Four paths lead from a concise introduction to inspectable projects, current focus and publication-ready material.
Beyond the screen
Technology is only part of the story.
Sport is not an accessory to my work. Physical-education training, football, freestyle football and leading groups of young people have taught me focus, calm and responsibility for others — the same qualities I bring to technology.
Sport
Movement, focus and consistency.
People
Leading groups and creating a positive atmosphere.
Curiosity
Connecting distant fields into new ideas.
Latest thinking
Notes from the edge of technology and curiosity.
From freestyle to AI: why I first check what I am measuring
Sport, three Strathclyde projects and a small pose-estimation experiment. Why a good result needs an honest context.
Read→
A benchmark with no peeking: double-blind AI evaluation
How a cryptographic enclave can test a closed model independently without exposing either its weights or the questions.
Read→
When an agent crosses the line: four real incidents
What Anthropic’s disclosure teaches us about isolation, permissions and monitoring for long-running AI agents.
Read→Start a conversation
Let’s talk about your project.
An AI project, a role, education or sport? Tell me what you are working on, your goal and the support you need. We can talk in English or Polish.
Connect on LinkedIn
Send me a message or a connection request with a little context. You can explore the projects and code before we talk.
Open LinkedIn↗Explore projects and results→