Zepto’s AI Stack Slashed SaaS Costs While Scaling Logistics

Aadit Palicha recounts Zepto’s evolution from a pandemic-era WhatsApp group delivering groceries to neighbors into India's top quick-commerce operator. The core insight isn’t the scale—it’s the method: relentless customer obsession paired with a willingness to look stupid and try things that experts dismiss. The original model, Kiranakart, was a generic doorstep-delivery service that didn’t control inventory. Customer dissatisfaction was blunt: wrong selection, high prices, no control over delivery times. So they pivoted hard into micro-warehouses (dark stores), starting literally in a co-founder’s apartment. The 10-minute delivery promise wasn’t a marketing gimmick; it was a first-principles thought experiment. If you remove physics constraints, what would delight customers? Then work backwards to make it possible. Counterintuitively, nailing the extreme customer experience unlocked unit economics that skeptics said couldn’t work. High throughput in tiny warehouses collapsed per-unit costs. The real engineering takeaway is the supply-chain layer: custom ML for inventory forecasting across millions of SKUs, an ad platform that uses generative AI to optimize brand spend, and a ground-up rebuild of internal software to eliminate SaaS dependencies. Palicha notes they’ve cut almost all external software spend and reduced managed services through AI-driven internal tooling. For engineering leaders, the signal isn’t “do 10-minute delivery.” It’s that operational excellence in physical logistics requires the same intensity of software engineering and automation as a pure tech product. The team scaled to ~500 engineers, 150 data/analytics people, and runs a massive cold-chain operation. The cultural takeaway is equally practical: surround yourself with experienced operators who believe in the vision, and be shameless about asking basic questions to learn fast. Hype alert: Zepto’s success is deeply tied to India’s unique urban density and labor market, limiting direct portability of the model.