Abstract

Buying and selling pre-owned vehicles in India remains riddled with trust deficits, unreliable pricing, and cum- bersome trade-in procedures. We built a web-based automotive marketplace that tackles these pain points head-on using three machine learning subsystems working in tandem. The first is a price estimation module powered by gradient boosted trees, trained on features we engineered specifically around Indian realities things like cumulative monsoon exposure, state-level supply-demand gaps, and the bureaucratic overhead of inter-state RTO transfers. The second component is a vehicle suggestion engine that blends neighborhood-based collaborative signals with attribute-level content matching, refined through a deep scoring network and a post-hoc re-ranker sensitive to budget limits, geographic taste, and catalog diversity. Third, an automated buy- back assessor fuses CNN-derived exterior condition grades with document verification scores to output a transparent trade-in quote. We validated every module against a curated pool of 152,437 real listings spanning all 28 Indian states. On price estimation our system lands at 6.8% MAPE, which is roughly a quarter better than a vanilla XGBoost without our regional features. Recommendation quality sits at 0.847 nDCG, edging past standard wide-and-deep baselines by over four percentage points. The buy-back pricer hits 7.3% MAPE on a holdout of 5,000 verified dealer transactions. Everything runs on a horizontally scalable microservice backbone that comfortably handles thousands of concurrent sessions.