- The problem
- Analysing transactions one row at a time can miss suspicious relationships between accounts. A group of connected accounts may reveal a pattern that is not obvious from any single transaction.
- What I built
- I built graph-based pipelines that represent transactions as a network, learn node embeddings and combine anomaly detection with supervised classification. SHAP helps explain the signals behind a flag, while a Streamlit interface makes the results easier to review.
- What happened
- Anomaly recall improved by 20%, and experiments covered graphs with more than 1 million nodes. The project received the Best System Award.
Projects and practical experiments.
These projects cover graph-based fraud detection, explainable prediction, computer vision and the engineering work around machine learning systems. Each section describes the problem, the approach and the available results.
- The problem
- Loan default models can influence important decisions, so it is useful to understand which factors contribute to a prediction.
- What I built
- A classification model for loan default risk, with SHAP-based explanations that help inspect which factors influenced each prediction.
- The problem
- A health-related risk score is easier to review when the factors behind the prediction are visible.
- What I built
- An end-to-end prediction pipeline using Logistic Regression and SVM, with SHAP explanations to show which features influence each result.
- What happened
- The reported result is 0.88 AUC-ROC. The project received Best Project Proposal (2024).
This is an educational project. It is not a medical diagnostic tool.
- The problem
- Vision models are easy to demo and harder to make fast and reliable enough to use.
- What I built
- Prototypes for drowsiness detection, object tracking and medical image classification, using data augmentation and a focus on runtime efficiency.
- What I took from it
- This work helped me examine model behaviour beyond a single score and understand the limits of each approach.
- The goal
- Take one model from notebook to a dependable service: reproducible training, an API, packaging, automated tests, experiment tracking and basic monitoring.
- Where it stands
- This project is in progress. The plan is shown separately from completed work, and features will be added here as they are implemented and tested.