A few projects that show how I move from a question to a shipped, measurable result.
General technical depth: modeling, engineering, evaluation.
Benchmarked 10 CNN configurations (EfficientNetB0 vs. ResNet50) for pneumonia classification from chest X-rays. The best model hit 71% accuracy, 92% recall, and a 0.84 ROC-AUC, built as a triage-prioritization aid rather than a standalone diagnostic.
ResNet50 · 92% recall · 0.84 ROC-AUC
Compared four approaches (N-gram+TF-IDF, fine-tuned DistilBERT, two SBERT variants) for binary sentiment classification. Fine-tuned DistilBERT won with F1 ≈ 0.91, beating the N-gram baseline and SBERT+LR.
DistilBERT · F1 ≈ 0.91
Built a regression pipeline to forecast daily sales revenue using lag features and moving averages. Tuned Lasso Regression (alpha=100) was the best model, evaluated on MAE and RMSE, designed for retail inventory planning.
Lasso regression · retail demand forecasting
Applied work tied to business outcomes: conversion, ROI, decisions.
Explored a fashion retailer's shopper data to find what drives spend and repeat buying across demographics, product category, subscription status, and promo usage. Subscribers clustered among higher-frequency repeat buyers, while promo codes weren't moving order value.
Personalization for a fashion e-commerce platform
Analyzed a classic car dataset for a restoration and auction client, answering five business questions on fuel efficiency, horsepower leaders (Chrysler and Cadillac), and weight and MPG patterns by country of origin.
Insight-driven EDA for restoration & auction
Interactive Tableau dashboard analyzing bookings across 19 destinations: revenue by destination, seasonality, satisfaction by channel, lead time, trip type, and channel performance, all in one filterable self-service view.
Interactive BI for travel booking decisions
Designed and analyzed a checkout A/B test on ~21,000 simulated sessions, from pre-test power analysis to a two-proportion z-test with Wilson confidence intervals. A one-page checkout lifted conversion 25.1% (p = 0.0008), consistent across device segments.
25.1% conversion lift · p = 0.0008