GreenCart cuts last-mile delivery cost 34% with AI route optimization

Title

GreenCart cuts last-mile delivery cost 34% with AI route optimization

Content

GreenCart, a Bengaluru-based D2C grocery platform, replaced a broken static dispatch system with an AI-driven route optimization engine and, in 20 weeks, cut cost per delivery 34% (₹42 → ₹28), lifted metro-peak SLA adherence from 82% to 96%, and raised fleet utilization 38% (68% → 85% active delivery time) across 18 cities. The project team — CTO Meera Iyer, Head of Logistics Arjun Desai, and Data Science Lead Priya Nair — treated the dispatch engine as a bounded problem rather than a platform rebuild. As the case study frames it, "rather than embracing an open-ended platform rebuild, the team treated the project as a focused dispatch-engine replacement — an approach that kept the scope tight enough to hit the deadline and the risk manageable enough that leadership would not pull funding halfway through." By mid-2024 GreenCart was running 42 dark stores and 1,200 delivery partners processing roughly 85,000 orders per week, but the legacy Node.js + PostgreSQL dispatch system had not been updated since 2022. Static zone-based FIFO routing meant partners returned to hubs with empty bags 32% of the time, and Mumbai and Pune missed the 90-minute SLA on 18% of peak-hour orders; app-store satisfaction had slipped from 4.6 to 3.9 stars in six months. The build played out over four phases. Weeks 1–3 consolidated siloed Postgres, Firebase, and WhatsApp data into BigQuery, instrumented GPS streaming every 15 seconds, and pulled in Mumbai Traffic Police and Indian Meteorological Department feeds. Weeks 4–8 had Priya Nair build a capacitated-vehicle-routing-with-time-windows (CVRPTW) model on Google OR-Tools, weighted on cost per kilometer, time-window violation probability, and partner idle time; back-testing against two months of historical dispatch data showed the model would have cut fleet kilometers 27% and lifted SLA adherence 11 points. Weeks 9–16 ran a minimum-viable rollout in Coimbatore (8 dark stores, 120 partners) using a strangler-fig adapter layer where Kafka duplicated order events, the Python optimizer emitted route assignments, and a Node.js consumer wrote them back into the legacy Postgres. Weeks 17–20 expanded to Pune and Hyderabad, then Mumbai — where adding metro stations as valid transfer nodes in the road graph improved routing efficiency another 6%. The production stack is Apache Kafka for ingestion (windowed into 60-second batches), a Python OR-Tools service on GCP Cloud Run exposing POST /optimize (5 GB memory, 30-second timeout, 1.2-second mean solve), a Node.js assignment adapter for backward compatibility with the existing driver app, and a Grafana dashboard alerting on >10% deviation from the 7-day rolling average. GPS noise from dense urban areas was smoothed with a Kalman filter. Operationally, deliveries per partner per shift rose from 6.8 to 9.4, idle time fell from 42 to 19 minutes per shift, manual dispatch interventions dropped 88%, false-reassignment fell from 15% to under 2%, partner churn dropped 12%, and app-store ratings recovered to 4.5 stars. The CFO validated roughly ₹2.1 crore in annualized EBITDA at the existing order volume, and freed capacity let GreenCart enter three new tier-2 cities in Q2 2025 without adding logistics headcount. GreenCart has approved phase two: SKU-level demand forecasting to reduce substitution and improve dark-store inventory turns, reusing the Kafka + BigQuery + Grafana foundation.