Hello, I’m Tan Bo Sheng

Chatbot database design (Group Project)

As part of my module Designing & Developing Databases project, for one of the question which was quite open ended in nature, I proposed a chatbot concept designed to handle exceptional cases such turbulence compensation scenarios as described in our project brief.

The video showcases:

• The design process and key assumptions made to maximise customer satisfaction and experience.
• My approach to solving an abstract problem through structured analysis and creative thinking.
• This project highlights my problem-solving skills, analytical thinking, and ability to design solutions for real-world cases.

Note: The original video snippet from the submission is used for authenticity purpose

Credit Default Risk (Group Project)

Worked in a team of 5 from multiple backgrounds to build an end-to-end ML pipeline that predict's customer credit default. The project covered EDA, data cleaning, categorical encoding, scaling, class-imbalance handling, model selection with cross-validation, and justifications. Technologies used includes, Python, Jupyter Notebook, Pandas, Numpy, Matplotlib, Seaborn, 'PLotly Express & Plotly Graph Objects', Scikit-learn, XGBoost, imblearn, KMeans

How It Was Done

• Performed data cleaning, handling missing values, encoding categorical variables, and scaling features.
• Conducted correlation analysis and visualization to identify key predictors.
• Addressed severe class imbalance using SMOTE oversampling, random undersampling, and class-weight adjustments.
• Trained and compared multiple models including Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, XGBoost, and an ensemble Voting Classifier.
• Optimized models using cross-validation and hyperparameter tuning with GridSearchCV.
• Evaluated models using Accuracy, Precision, Recall, F1-score, and ROC-AUC, with additional explainability provided through SHAP values.

Result Achieved

• Baseline models achieved high accuracy (~92%) but failed to identify defaulters effectively.
• Class-weighted and SMOTE-based models significantly improved recall to around 65–71%, making the system far more effective for risk flagging.
• Ensemble approaches, particularly weighted XGBoost and Voting Classifier, offered the best balance between recall and precision.

Technologies Used

• Python, Jupyter Notebook
• pandas, numpy, matplotlib, seaborn, plotly
• scikit-learn, XGBoost, imbalanced-learn, SHAP