Hank Zhu

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

Going deeper into Databricks architecture through hands-on platform design Recently worked on designing a production-grade Lakehouse platform on Databricks, focused on scalability, reliability, and governed data access. Key work from the project: • Designed a Medallion architecture processing TB-scale batch + streaming data from S3 and event streams • Built Auto Loader + Structured Streaming pipelines handling incremental ingestion and schema evolution • Used Delta Lake features (ACID, schema enforcement, time travel) to reduce data inconsistencies and enable recovery from bad loads • Developed modular PySpark transformation layers to standardize business logic across domains • Implemented Delta Live Tables with data quality expectations to automatically quarantine bad records • Applied Unity Catalog for centralized governance across multiple schemas and teams • Optimized compute strategy by separating job clusters for ETL and SQL warehouses for BI, improving cost control • Served curated gold datasets to BI tools via Databricks SQL for analytics consumption Next focus: strengthening automated monitoring, lineage visibility, and performance tuning across streaming + batch workloads. #Dataengineering #Databricks #AzureDatafactory #pyspark #C2C #C2H #newopportunities #USITJobs #Hiring Randstad USA

26

Why MCP Matters for the Future of AI As LLMs move from demos to real-world systems, the biggest challenge isn’t model capability—it’s context, action, and integration at scale. In this blog, I break down Model Context Protocol (MCP): • Why prompt-based integrations don’t scale • How MCP solves the M×N integration problem • Why MCP is emerging as the operating system for AI agents If you’re building agentic systems, RAG pipelines, or enterprise AI, MCP is a concept worth understanding early. 🔗 Read the blog here: https://lnkd.in/gJpcAHdt #AI #LLM #MCP #AgenticAI #Architecture #FutureOfAI

4

🔥 Scaling Big Data Processing with Databricks 🚀 At both American Express and CVS Health, I leveraged Databricks to build scalable, collaborative data pipelines and analytics solutions. Here’s what I achieved: 🔹 ETL & ELT Workflows – Built end-to-end pipelines using PySpark and Scala 🔹 Delta Lake – Enabled ACID transactions, schema evolution, and time travel for healthcare claims data 🔹 MLflow Integration – Tracked experiments and managed ML model lifecycle seamlessly 🔹 Job Orchestration – Automated pipelines with Databricks Jobs & Airflow 🔹 Performance Tuning – Optimized Spark jobs with partitioning, caching, and broadcast joins 🔹 Visualization – Created quick insights directly in Databricks notebooks for financial portfolio analytics 🔹 Use Cases: Risk adjustment, claim anomaly detection, transaction reconciliation ✅ Databricks simplified collaboration between engineering and analytics teams while ensuring scalability and cost optimization. Are you using Databricks for big data processing or ML? Let’s exchange insights! #Databricks #DeltaLake #PySpark #BigData #CVSHealth #AmericanExpress #MLflow #Spark #DataEngineering #Analytics

11

A Thought Exercise : Context Engineering with LLMs is analogous to Feature Engineering in Classical ML What are the similatieties and differences ? Share in the thread #AImusings courtesy: Jim Hong :)

28

9 Comments

Tredence’s Agentic AI Playbook How do enterprises scale AI beyond pilots? Tredence launches its Agentic AI Playbook, making agents peers (not assistants) in workflows. A structured guide for AI maturity & real business value. #AI #EnterpriseAI #AgenticAI https://lnkd.in/gSWj25ey

5

A Seasoned #DataScience Professional (Retail) | Simplifying #Data #AI #ML through real retail use cases | Host of “Truth wit Tuesdays 🚀” (Goes live Every Tues 9:30 AM IST) Last time we spoke of #Datalineage and why is it important to have mind map for every organization imbibed in their SOP's. Today we speak of other aspect of Data governance called "#DataAudit trail" Lets understand the Why's Why is Data audit required in first place? 1.To ensure Truth , Transparency & Accountability in how data is handled within an org 2. Enables Data Scientists to work on reliable data , enable faster debugging, support reproducibility & build trust with Stakeholders 3. Business can make better decision, have faster issue resolution & spread trust in analytics to executives & on field How does it help? 1.Data integrity & Accuracy: Keeps a track of all changes, who made them , when & what was modified 2.Compliance & Regulatory: Keeps Highly sensitive data or not in check to pass through some industry -country specific regulations (Like GDPR, HIPAA, SOX, RBI etc) 3.Accountability : Helps Establish clear responsibilities and data actions (create, update, delete , share) 4.Operational & Business Continuity: Enabling rollback /recovery incase of error /fraud/corruption What impact does Data audit create? 1. Forms Strong backbone of Data lineage & Metadata Management 2. Simplifies process for organization to understand how data flows through systems & how it is transformed over time 3. Detects unauthorized check ins strengthening organizations data security posture In short : Data audit is like CCTV camera for your data , it records every change, when something goes wrong , you can replay and find truth. #DataIntegrity #DataDrivenLeadership #EnterpriseTransformation #DigitalTransformation #TruthwitTuesdays

14

AI is not only disrupting Search, it is probably disrupting all interfaces.. Recently, I’ve fully shifted from classic SQL editors in Databricks to conversational AI — Cursor+ Databricks MCP Honestly, it’s addictive: - Just describe your data question - The Agent lays out the goal, writes the Python code, gives you charts, even adds recommendations - You skip the mechanical work and jump to “insights” 🚀 It’s seductive—until it isn’t. ⚠️ Today, it made up fictional business cases ( screenshot) to exaggerate the analysis impact. If I’d trusted that output, I would’ve made a high-visibility mistake. Here’s how I’m handling it now: 🔍 Added Rules - always use real data, and real company documents as examples 🧐 Always force the bot to show sample data sources What are you doing to keep AI-driven mistakes from slipping into your work?

5