Kuldeep Singh Sidhu

Twin brothers from India built an AI platform that achieved $15M in ARR in just 90 days. Is this the Shopify-for-apps moment? Here’s is the full story Emergent, founded by twin brothers Mukund (ex-Dunzo CTO) and Madhav Jha (ex-Dropbox/Amazon infra), just raised $23M Series A led by Lightspeed (total $30M to date). Their pitch? A multi-agent ‘vibe coding’ platform that turns plain English prompts into full-stack apps: frontend, backend, auth, payments, deployment, all automated. And the traction is wild: 👉 YC copy in mid-2024: 700k users, $10M ARR in 2 months. 👉 By Sept 2025: over 1M people building 2M apps and $15M ARR in 90 days. Why does it matter? Because Emergent isn’t just “no-code.” It aims to be the Shopify for apps, where anyone can generate production-ready tools, deploy them instantly, and eventually monetise them. Think jewellery chains spinning up AI-pricing apps, SMBs ditching spreadsheets for onboarding tools, even healthcare patients building pain-management solutions. But here’s the catch 👇 Retention > creation. Generating millions of apps is flashy. Keeping them live, secure, and monetised is harder. Discovery and billing flows are still in their early stages. Quality & compliance. Building real apps means adhering to PCI compliance, implementing secure authentication, and maintaining safe data practices. Emergent claims in-house testing/security agents, but independent audits will matter. Competition is fierce. Replit, Builder(dot)ai, Rocket(dot)New, Cursor: Everyone is chasing “apps from text.” Emergent’s moat must be infrastructure, reliability, and a marketplace. Unit economics. Hosting and runtime LLM costs scale with the number of active users. Using cheaper models and caching helps, but the test is focused on gross margin at scale. The opportunity: The low-code/AI app market is forecast to hit hundreds of billions by 2030+. If Emergent nails reliability, discovery, and monetisation, it could be the default platform SMBs use to generate and sell apps in days, not months. Question for you: If you could build a full-stack app from a single prompt, what would you build first? ➕ Save this post and follow Zubin Kutar ⚡ to Learn & Grow 10x with AI

189

22 Comments

Quantization in Large Language Models (LLMs) As LLMs scale, intelligence is no longer the bottleneck — memory, latency, and cost are. Quantization turns out to be one of the most powerful (and misunderstood) techniques enabling real-world deployment of large models. In this blog, I break down: • Why LLMs are surprisingly quantization-friendly • INT8 vs INT4 vs NF4 — what actually works in practice • How modern architectures absorb low-bit noise • Where quantization fails and how to design around it Quantization is no longer an optimization step — it’s foundational infrastructure for production LLM systems. Link: https://lnkd.in/gqGje75G #LLMs #AIEngineering #Quantization #MachineLearning #GenAI #MLOps

9

2 Comments

In this 20-minute talk, PolyMage Labs founder Uday Reddy Bondhugula talks about why and how one should build effective compilers for AI programming frameworks: https://lnkd.in/gWUQNUsF

79

Is there truly a gap in performance between online and offline RL training for LLMs? Here’s what the research says… TL;DR: There is a clear performance gap between online and offline RL algorithms, especially in large-scale LLM training. However, this gap can be minimized by incorporating on-policy samples into offline RL algorithms, creating an iterative or semi-online approach. RL for LLMs: A few variants of RL (i.e., RLHF and RLVR) are used for training LLMs. These techniques differ in how they derive a reward, but they both operate by generating completions for a set of prompts, computing rewards for these completions, and using those rewards to update the LLM’s parameters via an RL optimizer. RL optimizers: We usually use an (online) policy gradient-based RL optimizer (e.g., REINFORCE, PPO, or GRPO) to derive the policy update for an LLM. PPO-based RL has been the de facto choice in the past, but PPO: - Is memory intensive due to storing four copies of the model (policy, reference, reward and critic / value models). - Tends to be unstable and requires a lot of domain knowledge to set up correctly. Plus, online RL is difficult due to the complexity of generating on-policy completions during training. Instead, we could use a simpler offline algorithm like DPO, as well as a semi-online algorithm like iterative DPO or rejection sampling. (1) OOD data: In [1], the authors compare DPO with online RL algorithms, finding a clear bias in DPO. When DPO encounters out-of-distribution (OOD) or unseen responses, we see in [1] that the resulting LLM becomes biased toward these responses and behaves unpredictably. In contrast, PPO handles OOD data more robustly and learns a correct policy. We can mitigate these issues with DPO by performing iterative / semi-online DPO, where we: 1. Collect new, on-policy completions from the current LLM. 2. Score them with a reward model to create preference pairs. 3. Run DPO on the new data. 4. Repeat. (2) On-policy + negative gradient: Authors in [2] identify two key properties of alignment algorithms that lead to competitive performance: 1. On-policy sampling. 2. Negative gradients. Online RL algorithms naturally contain on-policy sampling. We can also incorporate on-policy samples into offline or RL-free algorithms by including on-policy samples in the training data. Negative gradients refer to the idea of pushing down the probability of “bad” completions. This naturally happens in PPO-based RL and DPO, but not in maximum likelihood algorithms like SFT. (3) DPO vs. PPO at scale: Authors in [3] perform an extensive comparison between DPO and PPO-based RLHF with large-scale LLMs (i.e., 70B parameters), whereas prior work mostly focuses on smaller LLMs in experimental settings. In this work, we see a clear gap between the performance of DPO and PPO at scale. However, authors find that the quality of training data still has a bigger impact on LLM performance than algorithm choice (i.e., online vs. offline).

152

6 Comments

Introducing BharatMLStack by Meesho: India's Production-Ready ML Infrastructure Platform. Meesho is democratizing machine learning at scale with a comprehensive platform that's already powering million-scale deployments across India and beyond. 🎯 Why BharatMLStack? ✅ Handle 1M+ feature vector retrievals per second ✅ Sub-10ms latency for real-time decisions ✅ 99.99% uptime with auto-scaling ✅ Petabyte-scale feature processing ✅ Cloud-agnostic Kubernetes deployment 🔧 Key Components: Horizon Control Plane - Orchestrates your entire ML ecosystem Trufflebox UI - Modern web console for ML management Online Feature Store - Real-time feature serving at scale 🇮🇳 India-First, Global Standards Built with Indian market needs in mind while maintaining enterprise-grade reliability. The platform is battle-tested in production environments, serving organisations from startups to enterprises. ⚡ Anyone can get Started in Minutes With Docker setup, sample data, and comprehensive documentation, you can deploy your first ML pipeline today. Link for the repo in comment.

30

5 Comments

History has a pattern most people ignore. When two brothers commit to one mission, outsized things tend to happen. These two just put India on the global AI product map. Mukund Jha and Madhav Jha didn’t start Emergent to “democratize coding.” They started it because app building was fundamentally broken. Especially for people building real businesses, not weekend side projects. Most people think building apps is hard because they’re non-technical. That’s a convenient myth. The real problem is the process. Too many tools. Too many handoffs. Too many systems that work fine until real users show up. And that’s exactly when founders can’t afford failure. So instead of hype decks and loud launches, these two brothers did something rare. They went quiet. No buzz. No flashy demos. No “AI will change everything” noise. Just months of grinding through the system, removing friction where it actually breaks. What came out of that effort became Emergent. Now here’s the part that should make you pause 👇 In just 7 months: 5M+ builders 6M+ real apps powered $50M ARR Users across 190 countries That kind of growth doesn’t come from hype. It comes from reliability. Which is why serious investors noticed. Early backing from Google AI Futures Fund, Gossamer Ventures, and SoftBank, followed by a $70M Series B. Not because it sounded exciting. But because it didn’t break when real users arrived. While most AI app builders stop at prototypes, Emergent goes end-to-end: Backend. Frontend. Database. Cloud. Payments. Deployment. Production-ready from day one. So what does this mean for the Indian Tech & IT landscape? This is the real shift 👇 1️⃣ India moves from services to software leverage When app creation becomes this fast and reliable, Indian founders can own products and IP, not just deliver projects. 2️⃣ Pressure on traditional IT models Billing hours and large dev teams won’t disappear overnight, but their dominance will shrink. Outcome-driven software beats manpower-driven software. 3️⃣ Faster startup cycles Founders no longer wait months to validate ideas. They can go from idea → production → users in days, not quarters. 4️⃣ Talent expectations will change The new edge won’t be “knows React or Java.” It’ll be: can you design systems, workflows, and outcomes? 5️⃣ India becomes a serious global builder, not just an executor Tools like this quietly shift India’s role from “cost-efficient tech” → product-first innovation hub. This isn’t just a startup success story. It’s a signal. Software is shifting from: "writing code to describing outcomes" And the teams that win won’t be the loudest. They’ll be the ones whose products don’t fail at scale. 👇 Let’s open this up Do you think tools like Emergent will disrupt Indian IT services or force them to reinvent themselves faster than ever? #IndianTech #IndiaAI #ArtificialIntelligence #AIProducts #StartupIndia #ProductManagement #ProductManager #ProductStrategy #AIInProduct #TechLeadership

11

Before LLMs became mainstream, the groundwork was already in motion. Abhay Tandon talks about running experiments back in 2017 while working at Target. The approach was simple but powerful: Use NLP to scan massive datasets. Identify the data points that actually matter. Extract signal from noise. Then use NLG to auto-generate reports. It wasn’t hype. It was structured automation. The core idea, though, hasn’t changed: 🔺 Automation scales analysis 🔺 Context increases decision quality 🔺 Generative AI compounds structured thinking AI didn’t suddenly appear in 2022. The foundation was always there. The tools just got exponentially better. 👉 Subscribe for more AI, data, and founder insights: https://lnkd.in/g3YQwF6h Ayush Bothra Shruti Sharma #AI #GenerativeAI #NLP #LLM #DataAnalytics #Automation #FounderJourney #InnovatorsAndDisruptors

8

Why Most Indian D2C Experiments Lie Text Thread 🫠 Indian D2C teams run a lot of experiments. Most of them don’t teach anything. Not because experimentation is wrong. But because the system is noisy. A few quiet truths: Traffic is inconsistent. Devices are slow. Networks fluctuate. User intent shifts minute to minute. So when conversion moves, teams celebrate or panic. Even when nothing meaningful changed. The problem isn’t lack of data. It’s lack of signal. Real learning happens when: • Experiments run long enough to settle • Variants don’t break edge cases • Failures are detected early, not post-launch • Teams know why something worked, not just that it did In India, small frictions amplify fast. So false positives are expensive. The best D2C teams won’t run more experiments. They’ll run cleaner ones. They’ll reduce noise before chasing uplift. They’ll trust outcomes only after systems are stable. In chaotic markets, clarity compounds faster than speed. #IndianD2C #CRO #Experimentation #Troopod #AICRO

15

9 Comments

Success stories where data science meets business innovation. 🌟LLMs Stories from Meesho, Amazon, Meta, Educational & Legal Sectors  youtu.be/tRycrqFmZbU  🌟DeepSeek R1: youtu.be/CkIVqoMIxsw 🌟GenAi Applications from Uber, Yelp, Vimeo & Instacart youtu.be/9w8_volSm6w 🌟GenAI powers DoorDash's e-commerce, Uber's operations, & ByteDance's youtu.be/SPR_-GXrZ2E 🌟How Companies Utilise RAG youtu.be/1JFPindItz4  🌟Uber Eats: Debiasing Position Bias in Recsys youtu.be/ZCO75OuMRY0 🌟Pinterest Ranking: youtu.be/WQj8QtjC3gA 🌟Spotify Audiobook Personalization via GNNs youtu.be/7lCt5JD9VJA  🌟ML System to Combat Counterfeit Fraud in E-Commerce youtu.be/YQZBgvLB_EQ  🌟Detecting Buyer-side Returns Fraud: youtu.be/as4i1tUo0EA 🌟LinkedIn's Budget Pacing for Targeted Ads youtu.be/R4EZ92VJvSI  🌟Netflix's Calibrated Recommendations youtu.be/DOWXNrBpO4w  🌟Netflix's Unified Recommendation ML Model: youtu.be/OKmv9sUrvk8 🌟Evolution of Recommendation Systems | Instagram Pinterest Twitter youtu.be/lgoyJn7MsH8 🌟Multi-Armed Bandit Strategies: youtu.be/2A5f3GrX0dA  🌟Overcoming Biases for a Better Recommender System youtu.be/oGb_mIdO0tA 🌟Building Scalable Query-Item Two-Tower Model based Retrieval System youtu.be/o-pZk5R0TZg  🌟Meituan's Game-Changing Two-Tower Retrieval Recsys Model | KDD'21 youtu.be/UhpbTSbi3lI  🌟Linkedin's Deep-leaning based CTR Modeling: The Deep, Wide, and Shallow Towers Explained youtu.be/7l0HLYVFEuU 🌟Lessons from Instagram’s Cutting-Edge Recommendation System youtu.be/Myna6rnmCG8 🌟How Twitter recommends Tweets: Inside Twitter's Recommendation Algorithm 🐦 youtu.be/IhGq9jgcxFM 🌟How eBay Built a Next-Generation Language Model to Revolutionise Online Shopping 🛒 youtu.be/h51nbWr7feo 🌟Leveraging Pre-trained Bert to generate Entity Embeddings across domain 📚 youtu.be/v-0J7o-nDBE 🌟Uplift Modeling to Detect Causal Effect | Uplift Modeling & Causal Inference to Improve Instagram ✨youtu.be/rKzG0Ct_ReA 🌟Advertiser Recommendation Systems at Pinterest 📌youtu.be/KEiBJ0fal3Q 🌟Dynamic Pricing Strategy to effectively fill slots | E-commerce | Travel | Hospitality Domains 💰youtu.be/a_CXpnsvPa0 🌟 Train your own Product Embeddings | Embeddings across Domains | Training on 1.02 Billion rows on GPU 💡youtu.be/DN4S96oHRhE 🌟 Model-Based Account Recommendations @ Twitter youtu.be/Xqo8fwgjxW4 🌟 Approximate Nearest Neighbour and a popular Library Annoy & Product Quantizer 🔍youtu.be/DSQOrBTqmYA & youtu.be/50PNumB7s3U 🌟 PID Controller to add a layer of diversity to the recommender systems 🎛️ youtu.be/laTxgnzjfR0 #datatrek #datascience #machinelearning #statistics #deeplearning #ai

3

📢 New Paper Alert! Dr. Aravind Reddy, Principal Project Scientist at CeRAI, WSAI, IIT Madras, recently co-authored a paper that was accepted at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026, one of the leading computer vision conferences. This was a joint work with Dr. Konda Reddy Mopuri and Mr. Saumyaranjan Mohanty, Department of AI, IIT Hyderabad. The paper, titled “DiRe: Diversity-promoting Regularization for Dataset Condensation”, proposes a simple, widely applicable diversity regularizer that can be plugged into existing dataset condensation methods. They show that DiRe consistently improves both test accuracy and dataset diversity across multiple benchmarks (CIFAR-10/100, Tiny-ImageNet, ImageNet-1K) and condensation methods, using principled metrics such as Coverage, Vendi Score, and CLIP-based coverage. Conceptually, the work supports CeRAI’s goals in Responsible AI by: ✅ enabling more data- and energy-efficient training via better synthetic datasets, ✅ making dataset diversity a quantifiable property rather than a vague claim, and ✅ strengthening the use of privacy-friendly synthetic data that remains robust across multiple architectures. Paper link: https://lnkd.in/gk2q5tpW Link to the code for the paper: https://lnkd.in/g7BtMQvw Balaraman Ravindran | Aravind Reddy | Centre For Responsible AI (CeRAI) | Wadhwani School of Data Science and AI, IIT Madras | Konda Reddy Mopuri | Saumyaranjan Mohanty | Department of AI, IIT Hyderabad | Robert Bosch Centre for Data Science and Artificial Intelligence (RBCDSAI) | IBSE: Centre for Integrative Biology and Systems medicinE | Indian Institute of Technology, Madras

49

#AI, #machineleanring #llm #GenAI, #AgenticAI Indore Institute of Science and Technology, The LLM Family Tree: Open-Source vs Closed-Source (Who’s Building What?) Large Language Models are evolving fast — but not all LLMs are built the same. Some are open for innovation, others are closed for control & scale. Let’s break it down 👇 🔓 Open-Source LLMs (Community-Driven Power) Best for: research, customization, cost control, on-prem deployment 🧠 LLaMA / LLaMA-2 / LLaMA-3 — Meta 🧠 Mistral / Mixtral — Mistral AI 🧠 Falcon — Technology Innovation Institute (UAE) 🧠 BLOOM — BigScience (Hugging Face) 🧠 Qwen (open variants) — Alibaba 🧠 Phi (small models) — Microsoft Fine-tuning freedom Transparent architecture Ideal for Agentic AI, RAG, and enterprise-specific models 🔒 Closed-Source LLMs (Enterprise-Grade & Polished) Best for: reliability, scale, production readiness 🤖 GPT-4 / GPT-4o / o-series — OpenAI 🤖 Claude — Anthropic 🤖 Gemini — Google DeepMind 🤖 Cohere Command — Cohere 🤖 Amazon Titan — AWS Strong reasoning & alignment Optimized infrastructure

50

Emerging Stack for LLM-powered Product - Ideation [1/3] Sharing insights from Kuldeep Yadav's presentation on Large Language Models—Stack(LLM Stack). Dive into key learnings on building, deploying, and differentiating LLM-powered products in real-world applications. The secret sauce to successful LLM-powered products lies beyond the base model—it's about how you design, build, and continuously iterate your product lifecycle. From scoping real user needs & measurable business metrics, to prompt engineering and robust evaluation, every step defines differentiation. Remember, everyone accesses the same foundational models; your stack and methodology are the true competitive edges. Varun Aggarwal Shailendra Nath Jha Shubham Bansal Jitesh Bhasin Negrah S

13

H-1B Shockwave: Impact on Indian AI Startups The $100k H-1B fee presents a tough choice for Indian AI startups with global ambitions. While challenging, it might just be the push needed to double down on local talent and innovation, keeping brilliant minds home to build the next wave of Indian tech unicorns. #AIStartups #H1B

5

Wrote a blog explaining "LoRA Fine-tuning Efficiency Under Different Loss Functions." LoRA made fine-tuning cheap, but not always efficient. This post explores how loss functions like Cross-Entropy, Label Smoothing, and Focal Loss change convergence, stability, and generalization in lightweight LLMs. Read here: https://lnkd.in/gh_vm2kG

2

🚨 Claude Cowork Agent Shock — A Turning Point for IT & The Rise of Agentic AI 🤖 The recent launch of Anthropic’s Claude “Cowork” Agent has sent ripples across global markets. Indian IT stocks dropped up to 8%, investors panicked, and headlines started asking — “Will AI Agents replace traditional IT services?” But as an Agentic AI Developer, I see this moment not as fear… but as a massive technological shift. 🌍 What Just Happened? Claude Cowork is not a chatbot. It is a digital worker capable of: - Executing real business tasks - Automating workflows end-to-end - Writing, analyzing, and decision-making - Using tools, files, and enterprise systems - Performing multi-step autonomous operations This marks the transition from AI Assistant → Autonomous AI Agent. 📉 Why Markets Reacted The fear is simple: If AI Agents can do: - Software tasks - Documentation - Analysis - Workflow automation Then traditional billable-hour IT services may reduce. But history shows — every tech disruption creates more opportunities than it destroys. 🔥 The Real Opportunity (Agentic AI Era) We are entering the age of Agentic AI Systems where: - Software will not just respond → it will act - Apps will not just process → they will decide - Automation will not just follow rules → it will reason This shift is creating huge demand for: ✔ Agentic AI Engineers ✔ RAG & Knowledge Systems Experts ✔ Workflow Automation Architects ✔ AI + Cloud + Systems Developers ✔ Enterprise AI Integrators 🧠 From My Perspective as an Agentic AI Developer This is not the end of IT. This is the evolution of IT into Intelligent Systems Engineering. The future belongs to developers who can build: - Autonomous Agents - Multi-Agent Systems - Tool-Using AI - Long-running AI workflows - Enterprise AI copilots AI will not replace engineers. Engineers who understand AI will replace those who don’t. 🚀 What Companies Should Do Now - Shift from traditional automation → Agentic automation - Invest in AI-first architecture - Train teams in LLM + Agents + Workflow AI - Build internal AI operators before competitors do 🔮 Final Thought Claude Cowork didn’t just launch a product. It signaled the beginning of the Autonomous Enterprise. And we, as developers, are not spectators — We are the builders of this new intelligent world. #AgenticAI #Claude #Anthropic #AIRevolution #GenAI #AIAgents #FutureOfWork #AITech #Automation #LangChain #AIEngineering #DigitalTransformation #AIInnovation

23

7 Comments

🔎 Hiring Panel Lens [Series 19] Why Strong PMs Sound Junior In AI Interviews After multiple AI-focused interviews, one pattern is very clear: Strong PMs are getting downgraded… not because of capability — but framing. When asked about AI work, most candidates start with: ➡️the model ➡️the feature ➡️the implementation Hiring panels are listening for something else: 👉 What decision the system is responsible for 👉 What happens when that decision is wrong 👉 How you designed for that failure Example: Weak answer: “We built an LLM to automate support.” Stronger answer: “We needed to decide which support queries could be resolved autonomously without increasing escalation risk.” Same system. Very different signal . AI products are not model problems. They are decision systems under risk. Quick self-check: When you describe your AI work, do you start with: A) The model B) The feature C) The decision the system makes Comment DECISION. I’ll share how strong candidates structure this in interviews. #productmanagement #aiproductmanager #productleadership #artificialintelligence #careergrowth #techcareers #hiringpanellens

16

3 Comments

100K active users in just 7 days + 12K paid conversions in week 1. That's NOT normal for a fresh AI platform — especially one building sovereign tech in India. Atomesus AI launched Jan 12, 2026 and the numbers already scream product-market fit. Here's what stands out for builders & founders 👇 🔥 Unlimited free tier (chats + image gen) hooks mass adoption fast 🔥 ~12% paid conversion right out of the gate — rare in AI land (most wrappers stay under 5% even months in) 🔥 Android live now, iOS coming soon, daily pushes from an ISRO-rooted team 🔥 Sovereign roadmap: hybrid today, full proprietary model ahead → real data-in-India + better Indic language handling This proves a few things we obsess over in 2026: → Generous free access + natural upsell beats forced subscriptions → Local talent + sovereignty focus builds actual moats in emerging markets → Iteration velocity > starting with the biggest model As someone tracking India's deep-tech moves, this milestone hits different because the monetization speed is what VCs chase in AI-native bets right now. Have you tried the app yet? → www.atomesus.com or search "Atomesus" on Play Store What's the real unlock for Indian AI teams? -Beating globals on cost + speed? -Sovereign data edges in a world of export controls? -Or hitting 10%+ paid conversion this early? Curious what devs, founders, and product folks think — drop your honest take below. Tag someone grinding on Indic AI or sovereign infra. Let's talk execution realities. 🚀 #IndiaAI #SovereignAI #DeepTech #AIStartups #ProductLedGrowth #MakeInIndia #AtomesusAI

10

Too Much Thinking Can Break LLMs: Inverse Scaling in Test-Time Compute Recent advances in large language models (LLMs) have encouraged the idea that letting models “think longer” during inference usually improves their accuracy and robustness. Practices like chain-of-thought prompting, step-by-step explanations, and increasing “test-time compute” are now standard techniques in the field. However, the Anthropic-led study “Inverse Scaling in Test-Time Compute” delivers a compelling counterpoint: in many cases, longer reasoning traces can actively harm performance, not just make inference slower or more costly. The paper evaluates leading LLMs—including Anthropic Claude, OpenAI o-series, and several open-weight models—on custom benchmarks designed to induce overthinking. The results reveal a rich landscape of failure modes that are model-specific and challenge current assumptions about scale and reasoning. Full Analysis: https://lnkd.in/gQdf4SE3 Paper: https://lnkd.in/gytEPFAK Project: https://lnkd.in/gtbPebjf Code: https://lnkd.in/g-Buqk2H Video Analysis: https://lnkd.in/g6jGpQBC Anthropic

36

1 Comment