Turning data into
better business decisions.

Sales Pipeline Analytics

Challenge: Sales leaders lacked visibility into funnel performance and pipeline velocity across deal stages.

Approach: Built SQL pipeline analysis and interactive Streamlit dashboards tracking deal progression.

Key Insight: Analysis identified late-stage negotiation as the largest source of pipeline leakage.

Business Value: Improved visibility into pipeline health and revenue forecasting reliability.

Complaint Risk & Operational Insights

Challenge: Escalating customer complaints were driving operational support costs and account churn.

Approach: Applied NLP text classification and XGBoost predictive risk scoring on customer logs.

Key Insight: Analysis highlighted recurring operational friction points associated with complaint escalation.

Business Value: Enabled early intervention triggers to reduce operational escalation costs and improve retention.

Thin-File Borrower Risk Assessment

Challenge: Traditional credit scoring excluded 31.4% of creditworthy thin-file applicants with sparse history.

Approach: Built alternative credit scoring models using utility & payment history in Python/XGBoost with SHAP.

Key Insight: Recurring digital utility payment consistency strongly correlated with low default risk.

Business Value: Scenario analysis shows credit criteria can expand safely without increasing credit default risk.

Mortgage Regulatory Data Quality

Challenge: Regulatory reporting required 100% data integrity before loan register submission.

Approach: Developed automated SQL reconciliation scripts and pre-submission audit rules.

Key Insight: Formatting discrepancies between loan origination systems and reporting tables caused errors.

Business Value: Designed an automated reconciliation framework capable of validating reporting datasets before submission.

Multilingual NLP Dissertation Benchmark

Question: Does translating multilingual customer feedback reduce ML classification accuracy?

Methods: Compared Logistic Regression, Random Forest, Naive Bayes and multilingual BERT.

Finding: Native-language fine-tuning consistently outperformed machine translation pipelines.