AI-Driven Trial-to-Purchase Prediction Model for Operational Efficiency in Fashion Quick-Commerce: A Case Study on ZODOK

Fashion quick-commerce platforms face a critical operational challenge: predicting which trial orders will convert to purchases. This paper presents an AI-driven trial-to-purchase prediction model developed as a case study for ZODOK, a Pune-based fashion quick-commerce startup operating on a try-before-you-buy model. A stacking ensemble combining Random Forest, Gradient Boosting, XGBoost, LightGBM, and a Multilayer Perceptron is trained on a 5,000-record dataset augmented from 1,000 real ZODOK transactions. Thirty-one engineered features capture customer loyalty, purchase intent, demographics, channel engagement, and temporal patterns. The proposed model achieves 82.0% accuracy, 91.4% recall, and an F1-Score of 88.2%, outperforming single-model baselines by up to 10.8 percentage points. Segment analysis reveals conversion rates ranging from 32% (window shoppers) to 96.6% (repeat customers). Deployment of the model is projected to yield a 5-percentage-point uplift in conversion rate, 30% reduction in inventory carrying cost, 29% reduction in logistics cost per order, and an annual margin impact of INR 3,71,496. A Flask-based API and web dashboard provide real-time predictions with interpretable factor explanations, demonstrating practical deployability for fashion quick-commerce operations.