Nrupali Kulkarni

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

Most AI failures today aren’t model failures — they’re context failures. As LLMs get more powerful, performance no longer comes from clever prompts alone. It comes from context engineering: designing what the model sees before it thinks — instructions, memory, retrieved knowledge, tools, constraints, and signals. Prompt engineering asks better questions. Context engineering builds a better thinking environment. This shift is what separates impressive demos from reliable, production-grade AI — especially for agentic systems, RAG pipelines, and real-world decision engines. Sharing some thoughts and learnings in my latest write-up on Context Engineering. Would love to hear how others are approaching this layer in their AI systems. Access here: https://lnkd.in/gBfFzZqQ #AI #LLMs #ContextEngineering #AgenticAI #RAG #MCP #AppliedAI

7

4 Comments

Explore how developers and students leveraged Free Edition to build AI agents, create rich visualizations and dashboards, and explore data engineering concepts

1

The Data Science Summer PD Fridays continues! This week’s picks: 📊 3 Steps to Master Graphing with Dataspire — Perfect for grades 4–10 teachers who want to make better graphs with less effort and more impact. 🧬 Evolution Education Workshop with DataClassroom — Join teachers and scientists at UVA’s Mountain Lake Station in the beautiful Blue Ridge Mountains. 🔬 Montana STEM Summer Institute — Explore how Science, Tech, Engineering, and Math combine to power 21st-century learning. Get all the details and more summer PDs at: https://hubs.li/Q03mBqsh0

3

🚀 Why Spark + Delta Lake is the Future of Data Infrastructure After years of working with traditional SQL environments, I can confidently say: Spark and Delta Lake are game-changers for modern data platforms. Here's why every data professional should pay attention: 📊 The Numbers Don't Lie: • 10x better compression - Your storage costs just got slashed • 10x faster queries - Even with perfectly tuned indexes and partitions in traditional SQL • Massive impact at scale - The difference becomes dramatic with TB+ datasets 💡 Why This Matters: While the performance gains might seem minimal with smaller datasets (tens of GB), once you hit hundreds of GB or terabytes, the difference is absolutely transformative. Traditional SQL optimization techniques like clustering indexes and partition columns simply can't compete with the distributed processing power of Spark combined with Delta Lake's intelligent data management. 🌟 The Open Source Advantage: What makes this even better? Both technologies are open source with enterprise-grade implementations available through: • Microsoft Fabric • Databricks • And other cloud platforms 🎯 My Take: Unless a completely revolutionary framework emerges, Spark + Delta Lake will remain the gold standard for enterprise data infrastructure. The combination of performance, cost efficiency, and scalability is unmatched. For data professionals: Now is the time to invest in these skills. Whether you're building your personal portfolio or your company's data platform, mastering these technologies will set you apart in the competitive data landscape. What's your experience with Spark and Delta Lake? Have you seen similar performance improvements in your projects? #DataEngineering #Spark #DeltaLake #DataInfrastructure #BigData #DataPlatform #ModernDataStack #DataArchitecture

4

Nikhil D. breaks down why calculating an initial trend and seasonality is so crucial for the STL process. Without this first step, you're essentially asking a smoothing algorithm to find patterns in pure noise, leading to unreliable results.

9

One of the biggest trade-offs in production RAG systems is retrieval quality vs latency. You can improve results by adding: • reranking • larger candidate sets • richer context But every additional step adds latency. A 2–3 second response may be fine for research assistants. But it can break the experience in customer-facing applications. The best AI architectures are not the most sophisticated ones. They’re the ones that fit within the latency budget of the product and the domain. #AIEngineering #GenAI #RAG #MachineLearning #ArtificialIntelligence

12

3 Comments

🎯 If Popular Brands Were Data Science Models — A Satirical Forecasting Thread After 10+ years building ML systems across CPG, finance, and healthcare, I’ve realized something: Every model has a brand personality. Here’s my take — and yes, I’ve debugged models that think ice cream causes drowning 🍦💀 🥤 Coca-Cola = Linear Regression 127 years old. Still here. Like COBOL, but fizzy. 🍫 Ferrero = Ensemble Methods Toys in chocolate? Overfitting to childhood chaos since 1946. 🧴 Unilever = Random Forest One tree sells Dove. Another sells Axe. The forest is confused, but profitable. 🍕 Domino’s = Gradient Boosting Started with cardboard crust. Boosted into a tech company. 🥣 Kellogg’s = ARIMA "Add more sugar!" "That’s not how ARIMA works." 🧼 P&G = Deep Learning Tide Pods went viral. Even the neural nets are confused. 🍟 McDonald’s = K-Means You: “I’m not hungry.” Model: “But it’s 3pm and you’re sad.” 🚬 Philip Morris = SVM Still searching for the hyperplane between “legal” and “lethal.” 🧠 Your Model = Spurious Correlation Thinks ice cream causes drowning. Send help. Which one hits too close to home? Drop your own brand-model combo 👇 #DataScience #AIHumor #Forecasting #MLOps #LLMs #FMCG #LinkedInHumor #FMCG #WhyDoIDoThisUnileverUnileverTheUnileverTheUnileverThe Coca-Cola CompanyGeneral MillsFerreroProcter & GambleMondelēz International Ivan SvetunkovAlok Patel, PhDArvind Kumar, PhDValeriy Manokhin, PhD, MBA, CQFJack RodenbergLancaster Centre for Marketing Analytics and Forecasting

10

3 Comments

New in CODATA Data Science Journal: Scott C. Edmunds, Nicole Nogoy, Qing Lan, Hongfang Zhang, Yannan Fan, Hongling Zhou, Chris Armit Title: Integrating Machine Learning Standards in Disseminating Machine Learning Research https://lnkd.in/eF-kdVeB #codata #FAIRdata #OpenScience #datascience #opendata #datasciencej

7

As Databricks adoption accelerates, platform leaders face mounting pressure to deliver reliable performance, predictable costs, and scalable data operations. But with a fragmented vendor landscape and no one-size-fits-all approach, selecting the right observability solution can feel overwhelming. That's why we created the Databricks Data Observability Buyer's Guide - a clear, actionable framework to help you cut through the noise. 📥 Download your free copy today: https://lnkd.in/gayN9ui9 #Databricks #DataObservability #FinOps #AIOptimization #CloudCosts

4

AI won't replace analysts. But analysts who use AI as a thinking partner will replace those who don't. Three of the sharpest data people I know just launched a course to help you become that analyst. Sravya (Senior Manager, Data Science @ Grammarly), Hai Guan (Head of Data @ Ontra), and Shane Butler (Principal Data Scientist @ Ontra) created AI Analytics for Builders on Maven. 𝐖𝐡𝐚𝐭 𝐭𝐡𝐞 𝐜𝐨𝐮𝐫𝐬𝐞 𝐭𝐞𝐚𝐜𝐡𝐞𝐬: → Translate vague product questions into clear, testable hypotheses → Define primary, secondary, and guardrail metrics with AI → Break down KPI movements through segmentation and trend analysis → Pressure-test your assumptions and edge cases → Turn analysis into clear, actionable recommendations 𝐓𝐡𝐞 𝐝𝐞𝐚𝐥: 🔥 5 early bird spots at $180 (normally $1,200) 🎟️ Everyone else who applies gets 40% off ⏰ Applications close in 24 hours I've worked with these folks. They're sharp, practical, and actually care about teaching skills that stick. If you're in data, product, or analytics, this is worth a look. 👉 𝐀𝐩𝐩𝐥𝐲 𝐡𝐞𝐫𝐞: https://lnkd.in/g7AHwSYY ♻️ Repost to help someone in your network level up their analytics game

71

24 Comments

Daniel Brown is preparing tools to help students solve real challenges facing our city for our upcoming Fall Datathon! Public transportation is vital for major cities and the importance of finding ways to make sure it's available to residents is vital. We're seeing cities like Philadelphia and Chicago losing bus service. Here in NYC there's multiple initiatives to maintain bus service with growing populations and vehicle traffic. First we need data that's accessible to begin analyzing and proposing solutions. Make sure to read Daniel's post below and find out how you can unlock NYC's bus story!

1

You’re in a Walmart ML interview. Interviewer: “What prediction algorithms do you know and when would you use them?” You: “Regression, Random Forest.” If that’s your full answer, this post is for you This is what production ML forecasting actually looks like 1. Linear / Regularized Regression (OLS, Ridge, Lasso) Use: strong priors, interpretability, causal decomposition (pricing, MMM) Breaks: non-linear interactions Used by: Marketing science, economics, Bayesian MMMs 2. Bayesian Structural Time Series (BSTS) Use: intervention analysis, incrementality, counterfactuals Breaks: weak controls, short pre-periods Used by: Google, Airbnb (CausalImpact) 3. Gradient Boosting (XGBoost / LightGBM / CatBoost) Use: tabular data + exogenous drivers (price, promo, weather) Breaks: pure time dependence Used by: Walmart, Amazon, Uber (enterprise demand) 4. ARIMA / SARIMAX (Dynamic Regression) Use: seasonality + known exogenous signals Breaks: regime shifts Used by: retail ops, energy, store-level forecasts 5. Prophet-based Models (incl. Robyn MMM) Use: decomposable trends, holidays, scalable MMM Breaks: high interaction effects, real-time attribution Used by: Meta (Robyn), marketing analytics Prophet is a component, not the strategy. 6. LSTM / Temporal DL Use: non-linear temporal regimes Breaks: low data, explainability needs Used by: e-commerce, logistics 7. Temporal Transformers Use: long-range dependencies, multi-entity planning Breaks: sparse data, latency constraints Used by: Amazon-scale forecasting systems 8. Monte Carlo Simulation Use: risk, uncertainty, scenario planning Breaks: bad assumptions Used by: finance, supply chain 9. Ensembles Use: error-critical forecasts Breaks: blind stacking Used by: Netflix, hedge funds, large retailers The answer Walmart wants: Stat baselines first → ML for exogenous effects → DL for regime shifts → causal validation → ensembles for stability. That’s not theory. That’s production ML [ Difference between college viva and senior MLs answer ] Comment D2C-STACK if you want the D2C focussed forecasting use cases and respective algorithms. #ai #forecasting #datascience #algorithms #deeplearning #timeseries #retail #d2c #analytics

29

1 Comment

💡 Real Healthcare Databricks Interview Questions & My Answers (8 Yrs Experience) Sharing 5 real-world scenario-based questions I faced as a Senior Data Engineer in Healthcare 👇 These reflect real production challenges from handling PHI to building scalable Delta pipelines. 1️⃣ How do you ensure HIPAA compliance while processing data in Databricks? 🩺 Answer: “In healthcare projects, I always enforce PHI protection using column-level encryption and access control via Unity Catalog. All PII fields are tokenized before landing in Delta tables, and audit logs are maintained using Databricks audit logs + Azure Monitor. We also integrate with Azure Key Vault for credential rotation and secure secrets management.” 2️⃣ How do you optimize Delta tables storing EHR/Claims data with billions of rows? 📊 Answer: “I apply ZORDER on high-cardinality columns like patient_id or claim_id for faster lookups. Partitioning is done by encounter_date or claim_year to improve pruning. I’ve also used OPTIMIZE with VACPUM scheduling and Delta caching for repeated analytical workloads — achieving a 45% boost in query performance.” 3️⃣ How do you handle incremental updates from HL7/FHIR streaming feeds? ⚙️ Answer: “I built a streaming ingestion using Auto Loader + Structured Streaming. Each micro-batch merges data into the Delta table using MERGE INTO with CDC logic. Watermarking ensures late-arriving updates are captured without duplication. Schema drift is handled via mergeSchema = true since FHIR resources often evolve. 4️⃣ How do you manage schema evolution across Bronze–Silver–Gold layers? 🏗️ Answer: “In our Medallion Architecture, we validate schema consistency at each layer. Bronze ingests raw JSON, Silver applies type casting + deduplication, and Gold provides curated analytics. To manage evolving schemas, I version schemas in Delta logs and maintain a schema registry ensuring backward compatibility.” 5️⃣ How do you build a unified pipeline for batch claims data and real-time clinical events? 🚀 Answer: “I designed a hybrid ingestion framework in Databricks that handles both historical (batch) data and real-time (streaming) HL7 feeds. Batch loads initialize the Delta tables, while streaming updates are applied using foreachBatch. The architecture supports both CDC and append-only streams for lab results, vitals, and patient admissions.” These real-world approaches have helped me build secure, compliant, and high-performing healthcare data platforms using Databricks, Azure, and Delta Lake. #Databricks #HealthcareDataEngineering #Azure #DeltaLake #MedallionArchitecture #CDC #HL7 #FHIR #DataEngineering #InterviewPreparation Kforce Inc IntraEdge Cardinal Health UnitedHealth Group Lakshya Technologies

10

Unlock the Power of Unified Data with Databricks In today’s data-driven world, organizations need more than just data storage they need data intelligence. That’s where Databricks shines bringing together data engineering, data science, and machine learning on one collaborative platform. Why Databricks Matters: 1. Unified Data Platform: Combines data lakes and warehouses for a “Lakehouse” architecture. 2. Scalable AI/ML: Enables rapid model training and deployment with built-in ML flow. 3. Collaborative Workspace: Data engineers, analysts, and scientists can work together seamlessly. 4. Open & Cloud-Ready: Built on Apache Spark, compatible with major cloud providers (AWS, Azure, GCP). At GrayCell America, we help organizations leverage Databricks to accelerate data transformation, AI adoption, and decision-making. From data silos to smart insights your data journey starts here. #Databricks #DataEngineering #AI #MachineLearning #DataScience #GraycellAmerica #DataAnalytics #DigitalTransformation

11

Databricks didn’t just release another AI feature. They made Custom Agents production-ready. That’s a big deal. Because most organizations aren’t struggling with ideas. They’re struggling with operationalization. The real question isn’t whether AI agents are powerful. It’s whether your data foundation is ready for them. Are you already building on Databricks, or still experimenting outside your core platform? #Databricks #AI #DataEngineering #Lakehouse #EnterpriseAI

4

New research from Databricks: LLMs Can Learn to Reason via Off-Policy RL Optimal Advantage-based Policy Optimization with Lagged Inference policy (OAPL) shows you don’t need strict on-policy training to improve reasoning. It matches or beats Group Relative Policy Optimization (GRPO), stays stable with large policy lag, and uses ~3× fewer training generations. For Databricks customers, it’s a simpler, practical, and equally powerful approach to RL that Databricks is pioneering internally — and bringing directly to Databricks customers, so enterprises can improve agents using the same methods we use for our in-house agents, without complex infrastructure changes. https://lnkd.in/gF5k49EK

53

2 Comments

📊 Tableau Desktop isn't just a tool — it's a career asset. After years of working with analysts and data teams, one truth stands out: professionals who master Tableau don't just build better charts. They make more informed decisions, tell impactful stories, and get noticed. We republished a deep-dive into what makes Tableau Desktop the industry standard in 2026 — and what's changed: ✅ AI-powered chart suggestions that cut visualization time by up to 70% ✅ 100+ native connectors from databases to cloud platforms ✅ Interactive dashboards that adapt to desktop, tablet, and mobile ✅ Built-in forecasting, clustering, and predictive analytics The platform keeps evolving. Are your skills keeping up? At Le Creative Lab, our Tableau courses are built for people who want to go beyond the basics — structured training, real-world workflows, and results you can show. 🎓 Explore our courses → lecreativelab.com/courses 📩 Get our articles straight to your inbox — subscribe to the Le Creative Lab blog on our website. Full article in the comments 👇 #Tableau #DataVisualization #DataAnalytics #BusinessIntelligence #DataSkills #LeCreativeLab #data #decisionmaking #PowerBI #AnalyticsTools

3

1 Comment