Meesho
National Institute of Technology Silchar
Bengaluru, Karnataka, India
2K followers
500+ connections
2K followers
500+ connections
About
I see Data as a product and actionable derived from this product as fuel to our business…
Activity
2K followers
-
Divay Jindal posted thisExcited to share that our paper “TrendPulse: A Simple yet Efficient Framework for Capturing Viral E-Commerce Spikes via LLM-Driven Contextualization” has been accepted at ACL 2026 (Industry Track) TrendPulse addresses the challenge of capturing short-lived demand spikes by detecting regional search momentum, transforming it into semantic trends using LLMs, and powering personalized recommendations via attention framework. This would not have been possible without an incredible team. Thanks to Devashish Gupta Arin Jain Bhavuk Singhal Also thankful for the support and guidance from Vinit Rongata Ravindra Yadav Debdoot Mukherjee Looking forward to presenting at ACL! #ACL2026 #MachineLearning #NLP #RecommenderSystems #LLM
-
Divay Jindal reposted thisDivay Jindal reposted thisHow do you personalize for users you know nothing about? At Meesho, we tackled the cold start problem head-on: blending demographics and early-session signals to deliver relevant recommendations from the very first scroll. Curious how we approached the cold start problem? 👉 https://lnkd.in/gfmg3b_b #Personalization #Ecommerce #MeeshoTech #MLForBharat Vinit Rongata Debdoot Mukherjee Ravindra Yadav Divay Jindal Devashish GuptaPersonalization from Day One: Solving the Cold Start Problem at MeeshoPersonalization from Day One: Solving the Cold Start Problem at Meesho
-
Divay Jindal shared thisGreat answerWhat skills do I need to be a data scientist at Google or Facebook?What skills do I need to be a data scientist at Google or Facebook?
-
Divay Jindal shared thisAwesome glossary. Have a look
-
Divay Jindal liked thisDivay Jindal liked thisAt the start of this year, I gave myself a challenge: write more regularly. Since I work at a payments company, I figured this was the most honest thing I could write about. Something I actually know. Or thought I knew. Thirteen weeks later, I realize how much I didn't. The comments, the DMs, the messages from people who work in banking, policy, and financial inclusion — they've taught me more than the research did. Someone pointed out last week that Ramesh, the migrant worker I wrote about in the KYC piece, could just open a Jan Dhan account in his home village. They were right. And that conversation pushed me to think harder about what "access" actually means when you're working six days a week in a different state. Two things hit harder than I expected when I actually sat down to write them out. The economics. I knew UPI was free. I hadn't thought long enough about what that actually means. The government subsidizes some of it. Banks absorb the rest. Nobody has made a permanent commitment. The most widely used payment system in the world's most populous country runs on a funding arrangement that gets renegotiated periodically. And KYC. I'd always filed it under compliance. Thirteen weeks of building toward it made me realize it's probably the single most consequential design decision in the entire stack. Everything else assumes the user already made it through the gate. For hundreds of millions of people, the gate is where the journey ends. The bigger realization though: none of these are isolated problems. Each layer of the system exists because the layer below it created a new problem. Money needs clearing. Clearing creates a single point of failure. Fixing that costs money. Free UPI creates fragile economics. And so on. Arc 1 is done. Thirteen weeks, one thread. Arc 2 starts next Tuesday. We go back to the early 2000s — before UPI, before smartphones, before any of this existed — and ask how it all got built. #HowMoneyMoves #Resolution
-
Divay Jindal liked thisDivay Jindal liked thisMost people use AI like a tool. Smart builders turn it into a system. That’s where 𝐒𝐊𝐈𝐋𝐋.𝐦𝐝 changes the game. It’s not just prompting. It’s about creating reusable AI capabilities 👇 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐒𝐊𝐈𝐋𝐋.𝐦𝐝? → A modular way to package AI tasks → Trigger-based execution → Works across tools like Cursor, Claude, Copilot 𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬: • Stop rewriting prompts again and again • Build reusable workflows • Standardize AI usage across teams • Scale AI like software 𝐖𝐡𝐚𝐭 𝐠𝐨𝐞𝐬 𝐢𝐧𝐬𝐢𝐝𝐞 𝐚 𝐒𝐊𝐈𝐋𝐋.𝐦𝐝? • Clear description (trigger logic) • Structured instructions • Constraints & rules • Tool integrations • Output format 𝐇𝐨𝐰 𝐢𝐭 𝐰𝐨𝐫𝐤𝐬: Input → Trigger match → Load SKILL.md → Execute instructions → Generate output 𝐁𝐞𝐬𝐭 𝐩𝐫𝐚𝐜𝐭𝐢𝐜𝐞𝐬: • Write as triggers, not prompts • Keep it modular • Add constraints for reliability • Avoid unnecessary complexity 𝐑𝐞𝐚𝐥 𝐢𝐦𝐩𝐚𝐜𝐭: Faster execution Consistent outputs Scalable AI workflows Most people are still prompting manually. The shift is happening towards 👉 𝐩𝐫𝐨𝐠𝐫𝐚𝐦𝐦𝐚𝐛𝐥𝐞 𝐀𝐈 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 P.S. Are you still writing prompts… or building reusable AI skills? Repost if this helped ♻️ —----- Looking for Gen AI Training or Automating your business processes using Gen AI? DM or comment "GenAI" to connect for 𝐅𝐑𝐄𝐄 Consultation Follow Manas Dasgupta for insights on Gen AI Automation
-
Divay Jindal liked thisDivay Jindal liked thisHaving worked with multimodal models for several years, one things I've faced far too often is that most multimodal models quietly throw away the signal they later struggle to recover. The issue is structural. Vision encoders are trained to produce compact features that align with text, not to preserve full scene information. That works for retrieval and captioning. It breaks for dense reasoning and generation. Once spatial detail is compressed, the rest of the system is reconstructing from a lossy prior. This new paper from MetaAI, Tuna-2, removes that failure point. It treats images as sequences of pixel patches and feeds them directly into a single transformer alongside text tokens. There is no CLIP style encoder, no VAE, no latent bottleneck. The same model handles perception and generation, operating on a shared token space. The implementation is heavier than it sounds. Pixel patches are embedded and interleaved with language tokens in long sequences. Training mixes objectives like next token prediction for text and pixel reconstruction or prediction for images. There is no stage where the model switches representations. Everything is optimized jointly. That forces the model to maintain fine grained visual structure while learning language alignment, instead of trading one off for the other. Whats impressive is that this holds up at scale. Encoder based systems usually win on efficiency and stability. Here, despite slower convergence and higher compute, the model surpasses them on fine grained perception and produces more consistent generations. That suggests the information loss from encoders is a real ceiling, not just an implementation detail. Compared to standard stacks, this is a different tradeoff. Less modular, harder to train, but cleaner in terms of information flow. You are not stitching together separately optimized components. My read is this meaningfully improves tasks that depend on exact grounding, where current models fail in subtle but important ways. If this direction continues, multimodal architectures simplify. The open question shifts from how to align modalities to how to efficiently train large models directly on raw sensory data. Always a pleasure to read papers from Lukes group!
-
Divay Jindal liked thisDivay Jindal liked thisRevolutionizing E-commerce Search: When LLMs Meet Hierarchical Taxonomies Researchers at eBay have developed a fascinating Chain-of-Thought approach that's transforming how e-commerce platforms categorize user queries. Instead of relying solely on traditional behavioral signals like click-through rates, this method leverages the semantic understanding of Large Language Models. How it works under the hood: The system treats query categorization as a tree traversal problem. Starting from the root of a product taxonomy, it uses an LLM to score the semantic relevance of each category branch on a 1-10 scale. The algorithm then applies dynamic thresholding- mapping scores to a standard normal distribution and pruning categories that fall below configurable selection and minimum thresholds. What makes this approach particularly clever is its breadth-first search strategy. At each taxonomy level, it only explores the most semantically promising branches, visiting just 1.7% to 24.8% of total category nodes while maintaining high accuracy. Key technical innovations: - Context-aware scoring that incorporates user intent (buying vs. browsing vs. seeking accessories) - Relative thresholding that adapts to score distributions at each node - Scalable variants using k-NN embedding filters for high-volume deployment - Built-in taxonomy diagnostics that identify structural gaps The results are impressive: Testing with models like Mixtral-8x7B showed 89.8% improvement in F1 scores over embedding-based baselines, with significant gains in both precision and recall across human judgment datasets. Beyond performance gains, this approach offers valuable insights for taxonomy optimization- automatically flagging categories like "Designer Sunglasses" that may be buried too deep in existing hierarchies. This research demonstrates how modern NLP can enhance traditional e-commerce systems while providing interpretable, actionable insights for platform optimization.
-
Divay Jindal liked thisDivay Jindal liked thisOne journey ends. Another begins. Mixed feelings, as always. As umami (formerly whatsyum) shuts operations, I'm excited to join Sahi as Head of AI. In just a couple of weeks, I already feel at home. The energy is real. The love for product and customer obsession shows up in every conversation. The pace of execution is surreal. The bar for talent is high. AI is a way of life. Grateful to Dale Vaz and Manish Jain for trusting me with what is perhaps one of the most transformational roles at Sahi. Lets build something great 🚀
-
Divay Jindal liked thisDivay Jindal liked thisIf Claude feels wrong. You're probably the problem. I've been using Claude Code daily for months now. One thing became obvious fast: When Claude gives you inconsistent results it's not the model, it's mostly the setup. Most people install it, type a prompt, and get frustrated. They never touch the folder that controls everything. The .claude/ directory is where Claude learns how to behave in your project: memory, automation, safety, and delegation. Here's what I configured that changed everything: 𝗖𝗟𝗔𝗨𝗗𝗘. 𝗺𝗱 Stop leaving it empty. Define your stack, architecture, and conventions. Claude reads it first every session. Treat it like onboarding docs for a new engineer. 𝗖𝗟𝗔𝗨𝗗𝗘.𝗹𝗼𝗰𝗮𝗹. 𝗺𝗱 Keep your personal preferences out of the shared project file. Your workflow shouldn't break your teammates'. 𝗺𝗰𝗽.𝗷𝘀𝗼𝗻 Configure it once. GitHub, JIRA, Slack, and databases. All connected, version-controlled. No more re-explaining your tooling every session. .𝗰𝗹𝗮𝘂𝗱𝗲/𝘀𝗲𝘁𝘁𝗶𝗻𝗴𝘀.𝗷𝘀𝗼𝗻 PreToolUse and PostToolUse guardrails that run automatically. Validation, linting, blocking unsafe operations — caught before they ship. .𝗰𝗹𝗮𝘂𝗱𝗲/𝗰𝗼𝗺𝗺𝗮𝗻𝗱𝘀/ Slash commands for recurring workflows. One keystroke runs your entire review or deploy process. No more copy-pasting the same prompt. .𝗰𝗹𝗮𝘂𝗱𝗲/𝘀𝗸𝗶𝗹𝗹𝘀/ Reusable workflows that load only when needed. Testing patterns, deploy checklists, API conventions. Auto-triggered by context or invoked manually. .𝗰𝗹𝗮𝘂𝗱𝗲/𝗮𝗴𝗲𝗻𝘁𝘀/ Specialized sub-agents with isolated context. Code review, security, and docs. Each has its own scope, instead of a single overloaded conversation. The simple test: If you keep repeating yourself to Claude, that instruction belongs in your setup. What's in your .claude/ folder right now? We could help you fix that! Check this: www.gentes.ai --- Kudos to Brij kishore Pandey for the amazing infographic.
-
Divay Jindal liked thisDivay Jindal liked thisModel size was never the problem. Memory was. And someone from China 🇨🇳 just shrunk it by 93%. 🎊 Anyone who's tried to actually use a long context window in production knows the real bottleneck. It's not the model. It's the KV cache. Every token your model processes stores key-value pairs in memory. As context length grows, that memory footprint grows with it. At a million tokens, the KV cache alone can consume more GPU memory than the model weights. That's why "supports 1M context" has been a spec sheet number, not a production reality. 🔧 DeepSeek V4 just made it real. Their hybrid attention architecture — Compressed Sparse Attention interleaved with Heavily Compressed Attention — fundamentally changes how tokens track long-range dependencies. 📊 The numbers: 10% of the KV cache footprint. 27% of the inference FLOPs. Same million-token context window. Flash variant: just 7% of the original cache size. 💰 The pricing reflects the efficiency: $0.14 per million input tokens for Flash. That's roughly 100x cheaper than comparable frontier APIs. 📌 The signal is clear. The next wave of inference cost reduction isn't coming from better models. It's coming from better memory architectures.
Experience & Education
View Divay’s full experience
See their title, tenure and more.
or
Publications
-
Basic Computer Skills
-
It is a 250 pages book covering all the basics skills needed to be taught to people such as Microsoft Office , basic computer architecture , introduction data entry tools and many more useful tools aiming at the skill development of the people in need so that could generate employment for themselves . This is approved under "Kaushal Vikas Yojna".
Other authors -
Courses
-
Convolutional Neural Networks for Visual Recognition
CS231n
-
Linear Algebra
18.06 Gilbert Strang
-
Machine Learning
Abu Mostafa
-
Machine Learning
Andrew Ng
-
Probability And Statistics
STAT 110
-
RL Course by David Silver
Reinforcement learning
-
Statistical Rethinking
Bayesian Machine Learning
Organizations
-
GYANSAGAR
Co-Founder
- PresentGyansagar is a Non Profit Organisation which aims at the skill development of the needy people. We teach English,Mathematics and Science to about 200 students living in the villages near to the campus of NIT Silchar. We conduct computer classes aiming at developing basic computer skill amongst the students .
Websites
- Portfolio
-
https://divayjindal95.github.io
Other similar profiles
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More