<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Kaustabh Datta Choudhury on Medium]]></title>
        <description><![CDATA[Stories by Kaustabh Datta Choudhury on Medium]]></description>
        <link>https://medium.com/@kaustabh93?source=rss-385fc2e502c5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*2F-jyrnSJsah99WMrV_VXA.jpeg</url>
            <title>Stories by Kaustabh Datta Choudhury on Medium</title>
            <link>https://medium.com/@kaustabh93?source=rss-385fc2e502c5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 05:17:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kaustabh93/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Personalized Search Ranking: The Zepto Way]]></title>
            <link>https://blog.zepto.com/personalized-search-ranking-the-zepto-way-496d0d405b71?source=rss-385fc2e502c5------2</link>
            <guid isPermaLink="false">https://medium.com/p/496d0d405b71</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Kaustabh Datta Choudhury]]></dc:creator>
            <pubDate>Mon, 01 Sep 2025 10:46:17 GMT</pubDate>
            <atom:updated>2025-12-17T14:04:35.402Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PMiv-Az0DhSPI3pptD8NYQ.png" /></figure><p>Search personalization in e-commerce is not a one-time switch, it’s a journey. At its earliest stage, platforms rely on a single, <strong>one-size-fits-all ranker</strong> that treats every user the same. While this works at small scale, it quickly breaks down when serving a diverse customer base with very different shopping intents.</p><p>At the other end lies <strong>N=1 personalization</strong> , where every individual gets a unique ranking tuned to their shopping behavior. Tens and hundreds of millions of diverse users’ shopping and interactions are modelled to build unique experiences for each of them.</p><p>This blog covers <strong>cohort-based ranking</strong> which sits as a middle ground. By clustering users into segments like Entry, Mass, Mass-Premium, and Premium, we capture the diversity of user needs without fragmenting the system into unmanageable silos. Cohorts based rankers are broad enough to be data-rich and stable, yet segmented enough to deliver meaningfully different experiences.</p><p>This blog is about how we unfolded cohort-based ranking with a deep neural network, using a <strong>Mixture-of-Experts (MoE)</strong> layer built on top of rich multi-modal data. In a subsequent blog post, we will tell the story of our journey from there to N=1 personalization.</p><h3>The Cohort Ranking Challenge</h3><p>Our user base isn’t homogeneous. Entry users behave very differently from Premium users.</p><ul><li><strong>Entry &amp; Mass users</strong> are <strong>price- and promotion-sensitive</strong>. Their decisions are driven by discounts, bulk packs, and affordability. For them, surfacing the best deals quickly is what keeps engagement high.</li><li><strong>Premium users</strong> are <strong>quality- and variety-seeking</strong>. They care about niche, organic, artisanal, or specialty SKUs, and expect the ranking to highlight relevance even in the long tail — not just what’s cheapest.</li><li><strong>Mass-Premium users</strong> sit in between, seeking a <strong>balance of savings and discovery</strong>. They want to find quality products but also respond strongly to promotions, making their ranking needs more nuanced.</li></ul><p>A single ranking model struggles to satisfy users with such different needs. For instance, when a <strong>Premium user</strong> searches for <em>“organic almond butter,”</em> a generic model may push a discounted peanut butter pack higher in the list because it optimizes for overall clicks. Meanwhile, an <strong>Entry user</strong> searching for <em>“cooking oil”</em> might instead see a niche cold-pressed variant promoted, even though their primary intent is affordability.</p><p>In both cases, the model is chasing aggregate engagement, but missing what matters most to each cohort — leaving neither user truly satisfied.</p><p>Our solution: <strong>segment users into Entry, Mass, Mass-Premium, and Premium cohorts</strong> and optimize search rankings specifically for each.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tk4iu223Rh5BZVXqY_7cyA.png" /><figcaption>Cohort ranking in action</figcaption></figure><h3>The Architecture Behind Cohort Ranking</h3><p>Designing a ranking model that works equally well for four very different user cohorts is not just a modeling problem , it’s an architectural one. Our system combines embeddings, a deep network, and a Mixture-of-Experts (MoE) layer to explicitly capture cohort-specific behaviors while keeping inference real-time ready.</p><h3>1. Rich, Multi-Modal Inputs</h3><p>We start by encoding a wide variety of signals:</p><ul><li><strong>Categorical Features</strong> (e.g. product &amp; search term categories, brand, offer flags) are embedded into dense vectors.</li><li><strong>Continuous Features</strong> (e.g. price, discount percentage, frequency counts) are normalized to stabilize training.</li><li><strong>Search Term &amp; Product Embeddings</strong> capture semantic intent and product relevance in the same vector space.</li><li><strong>Cohort Feature</strong> is itself learned as an <strong>8-dimensional embedding</strong>, letting the model generalize beyond one-hot cohort labels and capture nuanced cohort similarities.</li></ul><p>This ensures that the network receives a compact but information-rich representation of both query context and user context.</p><h3>2. Fusion via Concatenation &amp; Deep Network</h3><p>All features are concatenated into a single dense vector and passed through a <strong>multi-layer perceptron (MLP)</strong>. We use <strong>GELU activations</strong> for smoother gradient flow, and <strong>L1 regularization</strong> to encourage sparsity — which in practice reduces overfitting and highlights the most discriminative features.</p><p>This layer acts as a universal encoder, learning a baseline representation of relevance before routing into specialized experts.</p><h3>3. The Mixture-of-Experts (MoE) Layer</h3><p>The heart of the architecture is the <strong>MoE Layer</strong>, which balances global knowledge with cohort specialization.</p><ul><li><strong>Experts</strong>: Each cohort (Entry, Mass, Mass-Premium, Premium) is mapped to a specialized expert network. These experts are free to learn different ranking heuristics: for example, Premium’s expert learns to prioritize long-tail assortment, while Entry’s expert leans into promotions and discounts.</li><li><strong>Gating Network</strong>: Instead of hard-wiring users to experts, a small gating network computes <strong>dynamic weights</strong> based on the cohort embedding and deep features. This enables flexibility: even within a cohort, the gate can learn subtle distributional differences.</li><li><strong>Expert Logits</strong>: The outputs of all experts are combined using the gate’s weights (dot product), producing a final ranking score that is both cohort-informed and contextually tuned.</li></ul><p>This MoE setup prevents the “average-case dilution” problem of generic models while avoiding the inefficiency of training and serving completely separate models per cohort.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RfufCSS2wjHkHrjpg_2meg.png" /></figure><h3>4. Real-Time Inference at Scale</h3><p>Architectural elegance means little if it can’t serve production traffic. To make this design real-time ready, we engineered:</p><ul><li><strong>Custom data loaders with GPU sharding</strong> for efficient training.</li><li><strong>Multi-GPU distributed training</strong> for faster convergence.</li><li><strong>TensorFlow Serving on our in-house ML Platform aka Cerebro</strong>, optimized for throughput and latency.</li><li><strong>MLFlow tracked experiments</strong> for version control.</li></ul><p>The result: <strong>200K+ requests per minute</strong> with <strong>&lt; 51 ms latency; </strong>proving that personalization at this depth can be production-grade at Zepto scale.</p><h3>Business Impact</h3><p>Rolling out cohort-based ranking delivered measurable impact across our core business metrics:</p><ul><li><strong>Order per Day (OPD)</strong> rose by <strong>60 bps</strong>, reflecting stronger engagement and repeatability in user journeys.</li><li><strong>Average Order Value (AOV)</strong> increased by <strong>138 bps</strong>, showing that users were finding and buying products better matched to their intent.</li><li><strong>Gross Profit per Order (GPPO)</strong> surged by <strong>185 bps</strong>, a signal of richer baskets and healthier unit economics.</li></ul><p>For users, this translated into a sharper, more relevant experience: <strong>Premium cohorts</strong> now see curated long-tail and specialty SKUs, while <strong>Entry and Mass cohorts</strong> continue to get promotion- and price-led results. The balance is no longer a trade-off; each segment is served on its own terms, without compromising the other.</p><h3>What’s Next</h3><p>Cohort ranking was our first step in personalization at scale. But it’s just the beginning.</p><p>How did we improve upon this cohort based ranker? <strong>True user-level personalization</strong>, where search results adapt not just to a cohort, but to the individual. Think dynamic, behavior-driven personalization layered on top of cohort intelligence.</p><p>Stay tuned, we will cover this in a blog post soon.</p><h3>Final Thoughts</h3><p>Search personalization in e-commerce isn’t about showing “good” results. It’s about showing the right results to the right<strong> </strong>users at the right time.</p><p>By bringing together deep learning, scalable engineering, and thoughtful cohort design, we’ve built a search ranker that doesn’t just keep up with user diversity ; it thrives on it.</p><p>At Zepto, this is how we think about personalization: <strong>a system that learns from the nuance of our customers, and delivers impact where it matters most.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=496d0d405b71" width="1" height="1" alt=""><hr><p><a href="https://blog.zepto.com/personalized-search-ranking-the-zepto-way-496d0d405b71">Personalized Search Ranking: The Zepto Way</a> was originally published in <a href="https://blog.zepto.com/">Zepto TechXPress</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ironman 70.3 Warsaw: Almost There.]]></title>
            <link>https://medium.com/@kaustabh93/ironman-70-3-warsaw-almost-there-54a830413470?source=rss-385fc2e502c5------2</link>
            <guid isPermaLink="false">https://medium.com/p/54a830413470</guid>
            <dc:creator><![CDATA[Kaustabh Datta Choudhury]]></dc:creator>
            <pubDate>Fri, 20 Jun 2025 11:09:53 GMT</pubDate>
            <atom:updated>2025-06-20T11:09:53.003Z</atom:updated>
            <content:encoded><![CDATA[<p>Some races are about numbers. Others are about digging deep. This one was both.</p><p>When I signed up for Ironman 70.3 Warsaw, I had one clear goal: go sub-6. Training had gone well — I was fitter, stronger, and better prepared than ever before. Weekly hours hovered around 13, with long bricks, interval work, and structured recovery all falling into place. Add to that a renewed focus on strength training and nutrition (massive thanks to Strength Genesys), and I felt ready.</p><p><strong>Race day arrived crisp and cool — a rare treat in June.</strong></p><h3>🏊‍♂️ The Swim — 42 minutes</h3><p>The lake looked calm. I waded in, grateful for the gentle start — a long run-in helped ward off the usual cold shock (water temp was around 19.5°C). I felt smooth and strong. I found a good rhythm, drafted when I could, and sighted cleanly — no zigzagging this time.</p><p>Around midway though, I took a rogue goggle hit from a breaststroker that blurred vision in one eye. Not ideal. That fog stuck with me into T1 and beyond, but I didn’t let it mess with my momentum.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VZkhkG_sEoIEdL-I.JPEG" /></figure><h3>🚴‍♂️ The Bike — Avg ~31.5 km/h</h3><p>This was where things clicked.</p><p>The course was mostly flat but threw headwinds at us during some highway climbs. Still, I held steady — stayed aero, stayed disciplined. Nutrition and hydration went exactly as planned: gels and carb mixes in bottles.</p><p>The bike almost didn’t make it, though. While flying from India, the rear derailleur went out of alignment due to a shifter dislodging from the housing. Naveen from Cycle Station came through big time before I flew, and a local bike shop in Warsaw rescued the setup. Race morning, after a night of rain, I was just thankful I’d carried wet lube.</p><p>No mechanicals. No drama. Just execution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VnIoQiQu6PR3I5Sl.JPEG" /></figure><h3>🏃‍♂️ The Run — 2h 40m</h3><p>First km was chill. Legs were waking up. I built into it and by km 3, I was hitting target pace — hovering around 5:35/km. That’s when I took an on-course electrolyte.</p><p>Big mistake.</p><p>Within the next 2km, my stomach revolted. Severe cramps kicked in, and my pace went out the window. I had to walk stretches, duck into portaloos thrice, and abandon all pace goals. Switched to just water, gels, and salt tabs — no more aid station drinks.</p><p>There were moments where I genuinely wanted to DNF. But I couldn’t. Wouldn’t. I’d come too far. So I kept moving — slow, steady, hurting, but moving.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_eHEaC2WyS3lAV8G.JPEG" /></figure><h3>🏁 Finish Time: 6h 33m</h3><p>No, I didn’t break sub-6.<br>But yes, this was my fastest Ironman yet.</p><p>And more importantly, I walked away from the race not broken — unlike Westfriesland, where I could barely move post-race, this time I was up for a hike the very next day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oWSI1xGxoNMNVSag.JPEG" /></figure><h3>🙌 Gratitude</h3><ul><li>To Coach Sridhar, for the structured build-up that set me up for this performance.</li><li>To swim coach Nisha, for the stroke work that gave me a calm and confident swim.</li><li>To Strength Genesys, for the consistent, intelligent plans that elevated my conditioning.</li><li>And to Cycle Station, for always keeping my bike race-day ready.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sV1YDki-Esi91GMF.JPEG" /></figure><h3>🧠 Reflections</h3><p>What did I learn?</p><ul><li>Aid station electrolytes are not to be trusted — I’ll control my own nutrition next time.</li><li>Mental grit can carry you through where the body falters.</li><li>Sub-6 isn’t far. It’s right there. I’ve touched the edges of it. And I’ll be back for it.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*B0f1bqPNs0KVDc-e.JPEG" /></figure><p>Onward. Stronger. Smarter. Closer.</p><p>‍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=54a830413470" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lost in Translation? How We Fix Misspelled Multilingual Queries with LLMs]]></title>
            <link>https://blog.zepto.com/lost-in-translation-how-we-fix-misspelled-multilingual-queries-with-llms-173ce00c2ba1?source=rss-385fc2e502c5------2</link>
            <guid isPermaLink="false">https://medium.com/p/173ce00c2ba1</guid>
            <dc:creator><![CDATA[Kaustabh Datta Choudhury]]></dc:creator>
            <pubDate>Fri, 20 Jun 2025 08:47:21 GMT</pubDate>
            <atom:updated>2026-05-14T09:32:14.150Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RDDEsDko1HO6WAvW10afLw.png" /></figure><p>The rise of Large Language Models (LLMs) has generated significant excitement in the tech community. However, the real value lies in turning this innovation into practical, high-impact applications — especially in fast-paced, user-facing environments like quick commerce.</p><p>At Zepto, where a majority of our business is driven by in-app search, every query matters. A misspelled or poorly understood query can mean the difference between a successful purchase and a dropped session. In a landscape where users type in a mix of language &amp; script variants — often phonetically or on the go — getting query understanding right is not just a nice-to-have; it’s mission-critical.</p><p>This blog dives into one such deceptively simple but deeply impactful problem: detecting and correcting misspelled multilingual queries, particularly those written in English. We will walk through how we built a robust spell correction system using LLMs — tailored to the nuances of our user base and designed to meaningfully improve search relevance, user experience, and ultimately, conversion.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XxW3RSwQmCNJxMb_" /></figure><blockquote>With query correction, relevance skyrocketed — from 1 in 4 results being eggs to 4 out of 4.</blockquote><h3>The Core Challenge</h3><p>The task of detecting and correcting multilingual misspellings is complicated by the fact that many of our users input queries in vernacular languages using English lettering (e.g., typing “kothimbir” for “coriander”, “paal” which stands for “milk” in tamil). Most existing language models are built around native scripts, making it difficult to accurately understand and correct such queries.</p><h3>Our Approach: Building an MVP with Llama3</h3><p>We began by prototyping a multilingual spell corrector using API-based LLM playgrounds. After testing multiple models, we selected Meta’s <strong>Llama3–8B</strong> model, which provided a good balance of performance and accuracy for our use case.</p><p>The next step was to scale this solution. Relying on external APIs posed two major challenges: cost and reliability. To mitigate this, we hosted the Llama3 model on in-house Databricks’ model serving layer, which allowed us to integrate it with Spark jobs, ensuring high throughput and parallel access.</p><h3>Fine-Tuning for Accuracy</h3><p>After deploying the model, we turned to instruct fine-tuning — a lightweight yet powerful alternative to full model fine-tuning. Instead of modifying model weights, we focused on prompt engineering and system instruction design to teach the model specific behaviors like spelling correction, vernacular normalization, and disambiguation of mixed-language inputs.</p><p>We experimented with:</p><ul><li><strong>Role-specific system messages</strong>: We explicitly instructed the LLM to behave as a multilingual spell corrector, helping it narrow its focus and reduce hallucinations. For instance, we defined its role as “You are a spell corrector that understands multilingual inputs.”</li><li><strong>Few-shot examples</strong>: We included a diverse set of input-output pairs across English and vernacular queries (e.g., “skool bag” → “school bag”, “kothimbir” → “coriander leaves”) to demonstrate expected behavior, enabling the model to generalize better to unseen inputs.</li><li><strong>Stepwise prompting</strong>: Instead of asking the model to directly return the final corrected query, we broke down the process into intermediate steps like: (1) detect incorrect words, (2) correct them, and (3) translate if needed. This decomposition improved the model’s accuracy and transparency.</li><li><strong>Structured JSON outputs</strong>: To make downstream integration seamless, we enforced a consistent output schema — e.g., { “original”: “kele chips”, “corrected”: “banana chips” }. This reduced parsing errors and ensured clean handoff to our ranking and retrieval systems.</li></ul><p>This iterative instruct-tuning significantly improved task accuracy without increasing inference cost. It also allowed us to evolve the model quickly across use cases without retraining.</p><p>However, one issue that surfaced was the treatment of brand names. The model would often flag brand names as misspelled words. Initially, we included a list of brands in the prompt itself, but this increased the context length and inference time, reducing efficiency.</p><h3>Enter RAG: Retrieval Augmented Generation</h3><p>To overcome the limitations of static prompts and reduce dependency on long context windows. We implemented <strong>Retrieval Augmented Generation (RAG)</strong> — a paradigm that enhances LLM performance by grounding it with relevant, real-time contextual information from an external knowledge base.</p><h4>Our Architecture</h4><p>We designed a <strong>two-stage architecture</strong>:</p><ol><li><strong>Semantic Retrieval Layer<br></strong>Each incoming user query is converted into an embedding using a multilingual embedding model. This embedding is matched against product embeddings stored in a Vector DB<strong> </strong>using Approximate Nearest Neighbor (ANN) search.</li><li><strong>Contextual Prompt Construction<br></strong>The retrieved product results — including titles, brand names, and spelling variants (autosuggestion corpora) — are filtered, deduplicated, and compiled into a dynamically constructed prompt, which is passed to the LLM for correction.</li></ol><h4>Why This Matters</h4><p>Each component of this setup was built to address real-world challenges specific to a multilingual, quick commerce setting:</p><p><strong>1. Robustness to Noisy Inputs<br></strong>The retrieval engine operates in semantic space, so even severely misspelled or phonetically written queries return meaningful results.</p><ul><li><em>Example</em>:<br>Query: “<em>balekayi cheeps</em>”<br>→ Retrieved: <em>{ “title”: “Banana Chips”, “brand”: “Haldiram” }</em><br>→ Corrected: <em>“banana chips”</em></li><li><em>Example</em>:<br>Query: <em>“kottimbeer pudina”</em><br>→ Retrieved:<em> [“kothimbir (coriander)”, “pudina (mint)”]</em><br>→ Corrected: <em>“coriander and mint leaves”</em></li></ul><p>This helps us bypass brittle rule-based or edit-distance based approaches, which often fail in multilingual spelling variants.</p><p><strong>2. Brand Awareness<br></strong>Earlier, we tried stuffing the prompt with a hardcoded list of brands to prevent the model from “correcting” them. This bloated the prompt and degraded latency. With RAG, we inject only query-relevant brand terms.</p><ul><li><em>Example</em>:<br>Query: <em>“kellogs cornflex”</em><br>→ Retrieved: <em>{ “brand”: “Kellogg’s”, “product”: “Corn Flakes” }</em><br>→ Model output: <em>“Kellogg’s Corn Flakes”</em></li><li><em>Example</em>:<br>Query: “<em>valentaz</em>”<br>→ Retrieved: <em>{ “brand”: “Valentas”, “category”: “cardiac medication” }</em><br>→ Model confidently corrects: “<em>Valentas</em>”</li></ul><p><strong>3. Prompt Efficiency &amp; Latency Reduction<br></strong>By including only the top-k retrieved entries (e.g., top 5 product variants), we keep the token count low while injecting high-precision signals. This reduced prompt size by 30–40% compared to static dictionary-based prompts and lowered average inference latency by ~18s for a batch of 1,000 queries in our benchmarks.</p><ul><li><em>Before RAG</em>:<br>Prompt token count ~4,500+ tokens (with entire catalog dictionary &amp; brand list)</li><li><em>After RAG</em>:<br>Prompt token count ~1,200–1,400 tokens (only contextually relevant)</li></ul><p><strong>4. Dynamic Learning Capability<br></strong>RAG enables rapid adaptation as the catalog evolves. If a new product or brand enters the catalog, it’s automatically included in the recall without needing to retrain or reconfigure the prompt.</p><ul><li><em>Example</em>:<br>When we introduced “Sundrop Superlite Advanced Oil”, users began typing “sundrap oil” or “sundorp advanced”. These were dynamically grounded using vector retrieval without any manual intervention.</li></ul><h3>Outcome</h3><p>This RAG-enhanced setup transforms our LLM into a <strong>domain-aware, multilingual spell corrector</strong> that:</p><ul><li>Understands noisy, phonetically typed inputs</li><li>Differentiates brands from regular words</li><li>Adapts to evolving product catalogs</li><li>Delivers corrections with low latency and high precision</li></ul><p>This system is now a core enabler of our vernacular query understanding pipeline, and a blueprint we’re extending to adjacent use cases like voice-to-text query correction and personalized product explanations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/0*50WAnjglpEqwvCyJ" /></figure><h3>Leveraging User Reformulations as Implicit Feedback</h3><p>In parallel to the LLM-driven correction pipeline, we also incorporate user behavior signals to identify likely spelling corrections. Specifically, we monitor query reformulations within a short time window — typically within a few seconds of the initial search.</p><p>If a user quickly reformulates a query (e.g., from “banan chips” to “banana chips”) and the reformulated query shows a higher conversion rate, we treat it as a strong implicit signal that the original query was misspelled. These reformulation pairs help us:</p><ul><li>Auto-learn new misspelling variants,</li><li>Improve prompt coverage and examples,</li><li>Enrich training datasets for future supervised fine-tuning,</li><li>And reinforce the spell corrector’s decision-making in production.</li></ul><p>This behavior-driven feedback loop operates in the background and complements our LLM + RAG architecture with real-world user corrections — grounding the model in observed search behavior.</p><p>We have been building several other iterations to continuously improve the model, which we plan to cover in the future sequel to this blog.</p><h3>Results: Speed, Accuracy, and Scalability</h3><p>By fine-tuning our prompts and using RAG, we achieved a solution that was both fast and accurate. Hosting the model on Databricks ensured scalability, while instruct fine-tuning and RAG minimized costs and maintained high performance.</p><h3>Impact</h3><p>The implementation of this multilingual spell corrector had a significant positive impact on user experience and business metrics. By addressing incorrect, multilingual, and mixed grammar queries, we observed a <strong>7.5% increase in conversion rates </strong>for the impacted queries. This improvement highlights the critical role that accurate search query understanding plays in driving user engagement and overall performance in a quick-commerce setting. The project not only enhanced the search experience but also contributed to the bottom line by helping users find the right products more efficiently, leading to increased sales and customer satisfaction.</p><h3>Key Takeaways for Building with LLMs</h3><ol><li><strong>API Access vs. Self-Hosting</strong>: While API-based LLM services are great for prototyping, hosting models internally can provide better scalability and cost control in production.</li><li><strong>Instruct Fine-Tuning</strong>: Fine-tuning via instruct prompts, rather than the entire model, can save resources and improve task specificity.</li><li><strong>RAG for Efficiency</strong>: Retrieval Augmented Generation helps reduce context size and improve efficiency by providing only the necessary information to the model.</li><li><strong>Multi-Task Learning</strong>: LLMs can handle multiple tasks with careful prompt engineering, breaking down complex tasks into smaller, more manageable steps.</li></ol><p>At Zepto, LLMs are transforming how we solve real-world, complex problems. Our multilingual spell corrector is just the beginning, and we’re excited about what’s next.</p><p>Once we understand what the user meant, the next challenge is retrieving the right products for that intent, even when exact keywords do not match. That is the problem we explore in <a href="https://blog.zeptonow.com/how-we-built-high-precision-low-latency-semantic-search-in-production-75a6c61dee25"><strong>How We Built High-Precision, Low-Latency Semantic Search in Production</strong>.</a></p><p>Together, these systems form the first two layers of the search stack:</p><p><strong>Understand the query. Retrieve the right candidates. Then rank them intelligently.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=173ce00c2ba1" width="1" height="1" alt=""><hr><p><a href="https://blog.zepto.com/lost-in-translation-how-we-fix-misspelled-multilingual-queries-with-llms-173ce00c2ba1">Lost in Translation? How We Fix Misspelled Multilingual Queries with LLMs</a> was originally published in <a href="https://blog.zepto.com/">Zepto TechXPress</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>