Zepto

Zepto started as a quick-commerce platform for groceries. Today it’s where people shop for electronics, home essentials, gifts and festive needs. The catalogue has grown dramatically, but the app is still the same screen. That same real estate now needs to serve a far wider range of intents, in real-time, to millions of users a day. This is what makes in-session personalisation critical. A user browsing at 8 AM on a Tuesday (rushing to add staples before leaving for work) is operating very differently from the same user at 11:30 PM on a Friday, exploring snacks and beverages after a long week. While the historical user profile is identical, the immediate intent has completely shifted. To keep the shopping experience seamless, our ranking systems must instantly adapt to these context switches. In this post, we describe the system we built to do exactly that: a real-time dual sequence ranker that combines what a user is doing right now with what they have done before and serves scores wi

Material Event filed 2026-06-22

Much like our journey described in Debezium at Scale , our architecture relies heavily on real-time data flow. To understand user journeys, track operational metrics, and power our growth, we built Lucid — Zepto’s completely in-house product analytics engine designed to replace Mixpanel. Lucid captures millions of events per minute, routes them through Kafka, and dumps them into ClickHouse to give us lightning-fast, high-precision insights without the third-party SaaS pricing trap. We use Confluent Cloud to manage our Kafka infrastructure and the in-house ClickHouse Sink Connector. It was seamless — until our scale broke the default physics. Every hyper-growth engineering team eventually hits a wall where managed abstractions turn from a blessing into a bottleneck. For us, that wall appeared right at the intersection of Apache Kafka and ClickHouse. To ingest billions of events into ClickHouse for Lucid at a sustained throughput of 10 MB/s (peaking up to 15–20 MB/s), we hit a wall with

Every time a user types into a search box, they’re making an implicit bet — that the system on the other side understands not just what they typed, but what they meant, where they are, and what’s actually available to them right now. Most search systems get to ignore at least one of those dimensions. Zepto gets to ignore none of them. The query arrives. The clock starts, a structured chain of intelligence has to resolve intent — including attribute understanding, and query rewriting, fulfillment zone resolution, retrieval, ranking, all before the user lifts their finger. “ Algorithms rank documents, but great search understands people. ” Here is a search query: “amul doodh 1l”. It arrives from a device in Koramangala, Bengaluru at 11:43 AM on a Wednesday. In the next few hundred milliseconds — before the user lifts their finger — a structured chain of intelligence fires. The raw query passes through a correction layer that understands phonetically typed Indian vernacular. Blog on Query

Most teams find out their search is broken the same way: a customer complains, a business metric dips, or someone spots a glaring result in a demo. By then, the damage is done. The problem isn’t that search breaks. It’s that we have no reliable way to know when it does; or why. The tools most teams use to evaluate search quality are either too slow (human labelers), too indirect (click-through rate), or too narrow (head query benchmarks that ignore 80% of the query space). What follows is a framework built to fix that. It runs automatically, covers the full catalog, and produces a clean diagnostic signal that tells exactly where search is failing at the retrieval level, the ranking level, or both. Retrieval and ranking are different problems. Merging them is how teams spend months fixing the wrong thing. The two distinct ways search fails Before getting into the framework, it helps to be precise about what “bad search” actually means. There are two fundamentally different failure modes

At Zepto, the cart is more than a list of items. Behind every cart is an intent. Someone adding mozzarella, pasta, and basil isn’t randomly browsing; they’re making dinner. Someone with diapers, wipes, and formula is stocking up for their baby. Someone with chips, cola, and nachos is probably planning a night in. These patterns are obvious to a human observer. The challenge is teaching a system to see them too, and act on them in real time. The purpose of recommendations at Zepto isn’t just to show popular products. It’s to identify what a user is trying to accomplish as early as possible in their session and surface the items that help them get there faster. Instead of waiting for users to search for every ingredient individually, we want to proactively complete the picture. The cart contextual model is our solution to this problem. The Core Idea Every day, thousands of carts are created across the Zepto platform. Over time, strong patterns emerge. A large fraction of carts cluster ar

At Zepto, the cart is more than a list of items. Behind every cart is an intent. Someone adding mozzarella, pasta, and basil isn’t randomly browsing; they’re making dinner. Someone with diapers, wipes, and formula is stocking up for their baby. Someone with chips, cola, and nachos is probably planning a night in. These patterns are obvious to a human observer. The challenge is teaching a system to see them too, and act on them in real time. The purpose of recommendations at Zepto isn’t just to show popular products. It’s to identify what a user is trying to accomplish as early as possible in their session and surface the items that help them get there faster. Instead of waiting for users to search for every ingredient individually, we want to proactively complete the picture. The cart contextual model is our solution to this problem. The Core Idea Every day, thousands of carts are created across the Zepto platform. Over time, strong patterns emerge. A large fraction of carts cluster ar

At Zepto, speed isn’t just about delivery — it defines how we build, operate, and innovate across every layer of the stack. In large-scale, high-velocity supply chain operations like ours — where thousands of packers and riders work in tight coordination to enable 10-minute deliveries — even minor inefficiencies can cascade into operational delays. Traditionally, attendance tracking in such environments has relied on non-tech solutions like paper registers or basic digital systems such as app-based check-ins . These approaches often suffer from delays, manual errors, and, in many cases, identity-related frauds like proxy attendance or buddy punching. To overcome these challenges, we introduced face authentication-based attendance , a more comprehensive and tamper-proof solution. By verifying that the same person who was onboarded is the one marking attendance, the system not only enhances reliability but also protects the integrity of performance-linked incentives — such as those tied

At Zepto, speed isn’t just about delivery — it defines how we build, operate, and innovate across every layer of the stack. In large-scale, high-velocity supply chain operations like ours — where thousands of packers and riders work in tight coordination to enable 10-minute deliveries — even minor inefficiencies can cascade into operational delays. Traditionally, attendance tracking in such environments has relied on non-tech solutions like paper registers or basic digital systems such as app-based check-ins . These approaches often suffer from delays, manual errors, and, in many cases, identity-related frauds like proxy attendance or buddy punching. To overcome these challenges, we introduced face authentication-based attendance , a more comprehensive and tamper-proof solution. By verifying that the same person who was onboarded is the one marking attendance, the system not only enhances reliability but also protects the integrity of performance-linked incentives — such as those tied

Teaching a Tiny Model to Think Like a Giant When someone types “low fat milk” into Zepto, a race begins. Within milliseconds, our retrieval system pulls back around a hundred products that might match the query. Some are exactly right: “skimmed milk”, “2% fat milk”. Some are loosely related: “organic milk”, “toned milk”. And a few are simply enthusiastic guesses: “low-fat yogurt”, “protein shakes”. At this point, we already know something important: Search is not about finding products. Search is about ordering them. Because on a mobile screen, ordering is destiny. If the right milk is at position #1, the user taps and checks out. If it’s at position #14, they may never see it. And if the first result is wrong twice in a row, they might not come back. This is where the relevance ranker lives; in that thin, invisible layer between “results found” and “results trusted.” The Ten-Second Rule Imagine you walk into a grocery store and ask an employee for low-fat milk. They first hand you yog

Teaching a Tiny Model to Think Like a Giant When someone types “low fat milk” into Zepto, a race begins. Within milliseconds, our retrieval system pulls back around a hundred products that might match the query. Some are exactly right: “skimmed milk”, “2% fat milk”. Some are loosely related: “organic milk”, “toned milk”. And a few are simply enthusiastic guesses: “low-fat yogurt”, “protein shakes”. At this point, we already know something important: Search is not about finding products. Search is about ordering them. Because on a mobile screen, ordering is destiny. If the right milk is at position #1, the user taps and checks out. If it’s at position #14, they may never see it. And if the first result is wrong twice in a row, they might not come back. This is where the relevance ranker lives; in that thin, invisible layer between “results found” and “results trusted.” The Ten-Second Rule Imagine you walk into a grocery store and ask an employee for low-fat milk. They first hand you yog