Scalable System Design

We have seen recently a surge in vector databases in this era of generative AI. The idea behind vector databases is to index the data with vectors that relate to that data. Hierarchical Navigable Small World (HNSW) is one of the most efficient ways to build indexes for vector databases. The idea is to build a similarity graph and traverse that graph to find the nodes that are the closest to a query vector. Navigable Small World (NSW) is a process to build efficient graphs for search. We build a graph by adding vectors one after the others and connecting each new node to the most similar neighbors. When building the graph, we need to decide on a metric for similarity such that the search is optimized for the specific metric used to query items. Initially, when adding nodes, the density is low and the edges will tend to capture nodes that are far apart in similarity. Little by little, the density increases and the edges start to be shorter and shorter. As a consequence the graph is composed of long edges that allow us to traverse long distances in the graph, and short edges that capture closer neighbors. Because of it, we can quickly traverse the graph from one side to the other and look for nodes at a specific location in the vector space. When we want to find the nearest neighbor to a query vector, we initiate the search by starting at one node (i.e. node A in that case). Among its neighbors (D, G, C), we look for the closest node to the query (D). We iterate over that process until there are no closer neighbors to the query. Once we cannot move anymore, we found a close neighbor to the query. The search is approximate and the found node may not be the closest as the algorithm may be stuck in a local minima. The problem with NSW, is we spend a lot of iterations traversing the graph to arrive at the right node. The idea for Hierarchical Navigable Small World is to build multiple graph layers where each layer is less dense compared to the next. Each layer represents the same vector space, but not all vectors are added to the graph. Basically, we include a node in the graph at layer L with a probability P(L). We include all the nodes in the final layer (if we have N layers, we have P(N) = 1) and the probability gets smaller as we get toward the first layers. We have a higher chance of including a node in the following layer and we have P(L) < P(L + 1). The first layer allows us to traverse longer distances at each iteration where in the last layer, each iteration will tend to capture shorter distances. When we search for a node, we start first in layer 1 and go to the next layer if the NSW algorithm finds the closest neighbor in that layer. This allows us to find the approximate nearest neighbor in less iterations in average. ---- Find more similar content in my newsletter: TheAiEdge.io Next ML engineering Masterclass starting July 29th: MasterClass.TheAiEdge.io #machinelearning #datascience #artificialintelligence

Load Balancing: Beyond the Basics - 5 Methods Every Architect Should Consider The backbone of scalable systems isn't just about adding more servers - it's about intelligently directing traffic between them. After years of implementing different approaches, here are the key load balancing methods that consistently prove their worth: 1. Round Robin Simple doesn't mean ineffective. It's like a traffic cop giving equal time to each lane - predictable and fair. While great for identical servers, it needs tweaking when your infrastructure varies in capacity. 2. Least Connection Method This one's my favorite for dynamic workloads. It's like a smart queuing system that always points users to the least busy server. Perfect for when your user sessions vary significantly in duration and resource usage. 3. Weighted Response Time Think of it as your most responsive waiter getting more tables. By factoring in actual server performance rather than just connection counts, you get better real-world performance. Great for heterogeneous environments. 4. Resource-Based Distribution The new kid on the block, but gaining traction fast. By monitoring CPU, memory, and network load in real-time, it makes smarter decisions than traditional methods. Especially valuable in cloud environments where resources can vary. 5. Source IP Hash When session persistence matters, this is your go-to. Perfect for applications where maintaining user context is crucial, like e-commerce platforms or banking applications. The real art isn't in picking one method, but in knowing when to use each. Sometimes, the best approach is a hybrid solution that adapts to your traffic patterns. What challenges have you faced with load balancing in production? Would love to hear your real-world experiences!

I love it when AI works out, because when it does - it’s magic. Here is my personal 5-step readiness checklist so you succeed with it. 𝗦𝘁𝗲𝗽 𝟭: 𝗔𝘂𝗱𝗶𝘁 𝗬𝗼𝘂𝗿 𝗗𝗮𝘁𝗮 𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 Before any AI conversation, ask: "Is our data clean, accessible, and flowing properly?" - Map your current data sources and quality. - Identify gaps between systems. - Ensure data governance policies are in place 𝗦𝘁𝗲𝗽 𝟮: 𝗔𝘀𝘀𝗲𝘀𝘀 𝗬𝗼𝘂𝗿 𝗧𝗲𝗮𝗺'𝘀 𝗧𝗲𝗰𝗵𝗻𝗼𝗹𝗼𝗴𝘆 𝗖𝗼𝗺𝗳𝗼𝗿𝘁 𝗭𝗼𝗻𝗲 Meet your people where they are, not where you want them to be. - Evaluate current tool proficiency (Are they Excel natives? Advanced analytics users?) - Identify the skills gap between current state and AI requirements. - Plan bridge training programs. 𝗦𝘁𝗲𝗽 𝟯: 𝗕𝘂𝗶𝗹𝗱 𝗔𝗜 𝗟𝗶𝘁𝗲𝗿𝗮𝗰𝘆 𝗔𝗰𝗿𝗼𝘀𝘀 𝗬𝗼𝘂𝗿 𝗢𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Create understanding before implementation. - Run AI awareness sessions for leadership and end-users. - Define AI terminology and use cases relevant to your industry. - Address concerns and misconceptions upfront. 𝗦𝘁𝗲𝗽 𝟰: 𝗦𝘁𝗮𝗿𝘁 𝗦𝗺𝗮𝗹𝗹 𝘄𝗶𝘁𝗵 𝗣𝗶𝗹𝗼𝘁 𝗣𝗿𝗼𝗴𝗿𝗮𝗺𝘀 Test the waters before diving in. - Choose one high-impact, low-risk use case. - Select a team that's excited about innovation. - Measure adoption rates, not just performance metrics 𝗦𝘁𝗲𝗽 𝟱: 𝗘𝘀𝘁𝗮𝗯𝗹𝗶𝘀𝗵 𝗦𝘂𝗰𝗰𝗲𝘀𝘀 𝗠𝗲𝘁𝗿𝗶𝗰𝘀 𝗮𝗻𝗱 𝗙𝗲𝗲𝗱𝗯𝗮𝗰𝗸 𝗟𝗼𝗼𝗽𝘀 Define what winning looks like. - Set clear ROI expectations. - Create channels for user feedback and iteration. - Plan for scaling successful pilots Organizations that complete this readiness checklist see 3x higher adoption rates and significantly better long-term ROI. AI implementation isn't a sprint, it's a strategic marathon. Where is your organization in this readiness journey? What step are you focusing on right now?

‼️Ever wonder how data flows from collection to intelligent action? Here’s a clear breakdown of the full Data & AI Tech Stack from raw input to insight-driven automation. Whether you're a data engineer, analyst, or AI builder, understanding each layer is key to creating scalable, intelligent systems. Let’s walk through the stack step by step: 1. 🔹Data Sources Everything begins with data. Pull it from apps, sensors, APIs, CRMs, or logs. This raw data is the fuel of every AI system. 2. 🔹Ingestion Layer Tools like Kafka, Flume, or Fivetran collect and move data into your system in real time or batches. 3. 🔹Storage Layer Store structured and unstructured data using data lakes (e.g., S3, HDFS) or warehouses (e.g., Snowflake, BigQuery). 4. 🔹Processing Layer Use Spark, DBT, or Airflow to clean, transform, and prepare data for analysis and AI. 5. 🔹Data Orchestration Schedule, monitor, and manage pipelines. Tools like Prefect and Dagster ensure your workflows run reliably and on time. 6. 🔹Feature Store Reusable, real-time features are managed here. Tecton or Feast allows consistency between training and production. 7. 🔹AI/ML Layer Train and deploy models using platforms like SageMaker, Vertex AI, or open-source libraries like PyTorch and TensorFlow. 8. 🔹Vector DB + RAG Store embeddings and retrieve relevant chunks with tools like Pinecone or Weaviate for smart assistant queries using Retrieval-Augmented Generation (RAG). 9. 🔹AI Agents & Workflows Put it all together. Tools like LangChain, AutoGen, and Flowise help you build agents that reason, decide, and act autonomously. 🚀 Highly recommend becoming familiar this stack to help you go from data to decisions with confidence. 📌 Save this post as your go-to guide for designing modern, intelligent AI systems. #data #technology #artificialintelligence

𝘛𝘩𝘪𝘴 𝘸𝘢𝘴 𝘴𝘰𝘮𝘦𝘵𝘩𝘪𝘯𝘨 𝘐’𝘷𝘦 𝘣𝘦𝘦𝘯 𝘱𝘶𝘵𝘵𝘪𝘯𝘨 𝘵𝘰𝘨𝘦𝘵𝘩𝘦𝘳 𝘵𝘩𝘪𝘴 𝘸𝘦𝘦𝘬. 𝐍𝐨𝐭 𝐚𝐥𝐥 𝐀𝐈 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞𝐬 𝐚𝐫𝐞 𝐜𝐫𝐞𝐚𝐭𝐞𝐝 𝐞𝐪𝐮𝐚𝐥. Here’s how I integrate Microsoft Azure services to create AI that works for just about any business not the other way around. Want to know the secret sauce? 👇 7 Lessons from Building Scalable AI Solutions Customers Love: 𝐒𝐭𝐚𝐫𝐭 𝐰𝐢𝐭𝐡 𝐜𝐥𝐞𝐚𝐧 𝐝𝐚𝐭𝐚. ↳ Use 𝐀𝐳𝐮𝐫𝐞 𝐃𝐨𝐜𝐮𝐦𝐞𝐧𝐭 𝐀𝐧𝐚𝐥𝐲𝐳𝐞𝐫 for structured ingestion. ↳ Automate preprocessing with 𝐀𝐳𝐮𝐫𝐞 𝐅𝐮𝐧𝐜𝐭𝐢𝐨𝐧 𝐀𝐩𝐩𝐬. ↳ Store data securely in 𝐀𝐳𝐮𝐫𝐞 𝐁𝐥𝐨𝐛 𝐒𝐭𝐨𝐫𝐚𝐠𝐞. 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫 𝐟𝐞𝐚𝐭𝐮𝐫𝐞𝐬 𝐜𝐮𝐬𝐭𝐨𝐦𝐞𝐫𝐬 𝐯𝐚𝐥𝐮𝐞. ↳ Focus on actionable insights, not noise. ↳ Leverage 𝐀𝐳𝐮𝐫𝐞 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 for advanced prep. ↳ Collaborate with end users for relevance. 𝐓𝐫𝐚𝐢𝐧 𝐦𝐨𝐝𝐞𝐥𝐬 𝐭𝐡𝐚𝐭 𝐚𝐥𝐢𝐠𝐧 𝐰𝐢𝐭𝐡 𝐛𝐮𝐬𝐢𝐧𝐞𝐬𝐬 𝐠𝐨𝐚𝐥𝐬. ↳ Test multiple architectures, like custom LLMs. ↳ Use 𝐀𝐳𝐮𝐫𝐞 𝐌𝐋 and Azure OpenAI to streamline experimentation. ↳ Optimize for speed and scalability. 𝐃𝐞𝐩𝐥𝐨𝐲 𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐝𝐢𝐬𝐫𝐮𝐩𝐭𝐢𝐧𝐠 𝐰𝐨𝐫𝐤𝐟𝐥𝐨𝐰𝐬. ↳ Host on 𝐀𝐳𝐮𝐫𝐞 𝐊𝐮𝐛𝐞𝐫𝐧𝐞𝐭𝐞𝐬 for reliability. ↳ Use 𝐀𝐳𝐮𝐫𝐞 𝐅𝐮𝐧𝐜𝐭𝐢𝐨𝐧𝐬 for seamless integration. ↳ Monitor deployment with feedback loops. 𝐌𝐚𝐤𝐞 𝐝𝐚𝐭𝐚 𝐫𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐛𝐥𝐞 𝐚𝐧𝐝 𝐚𝐜𝐭𝐢𝐨𝐧𝐚𝐛𝐥𝐞. ↳ Index with 𝐀𝐳𝐮𝐫𝐞 𝐂𝐨𝐠𝐧𝐢𝐭𝐢𝐯𝐞 Search. ↳ Store outputs in 𝐂𝐨𝐬𝐦𝐨𝐬 𝐃𝐁 for scalability. ↳ Ensure query optimization for real-time use. 𝐁𝐫𝐢𝐝𝐠𝐞 𝐀𝐈 𝐰𝐢𝐭𝐡 𝐛𝐮𝐬𝐢𝐧𝐞𝐬𝐬 𝐥𝐨𝐠𝐢𝐜. ↳ Use 𝐀𝐳𝐮𝐫𝐞 𝐅𝐮𝐧𝐜𝐭𝐢𝐨𝐧𝐬 to support decisions. ↳ Automate workflows for better efficiency. ↳ Integrate insights directly into operations. 𝐆𝐨𝐯𝐞𝐫𝐧 𝐰𝐢𝐭𝐡 𝐬𝐞𝐜𝐮𝐫𝐢𝐭𝐲 𝐚𝐧𝐝 𝐚𝐠𝐢𝐥𝐢𝐭𝐲 𝐢𝐧 𝐦𝐢𝐧𝐝. ↳ Use 𝐆𝐢𝐭 𝐅𝐥𝐨𝐰 for version control. ↳ Secure pipelines with 𝐂𝐡𝐞𝐜𝐤𝐦𝐚𝐫𝐱. ↳ Automate infrastructure with 𝐓𝐞𝐫𝐫𝐚𝐟𝐨𝐫𝐦. Which step will move your business forward today? ♻️ Repost to your LinkedIn followers and follow Timothy Goebel for more actionable insights on AI and innovation. #ArtificialIntelligence #AzureCloud #InnovationInTech #AITransformation #MachineLearningPipeline

We’re entering an era where AI isn’t just answering questions — it’s starting to take action. From booking meetings to writing reports to managing systems, AI agents are slowly becoming the digital coworkers of tomorrow!!!! But building an AI agent that’s actually helpful — and scalable — is a whole different challenge. That’s why I created this 10-step roadmap for building scalable AI agents (2025 Edition) — to break it down clearly and practically. Here’s what it covers and why it matters: - Start with the right model Don’t just pick the most powerful LLM. Choose one that fits your use case — stable responses, good reasoning, and support for tools and APIs. - Teach the agent how to think Should it act quickly or pause and plan? Should it break tasks into steps? These choices define how reliable your agent will be. - Write clear instructions Just like onboarding a new hire, agents need structured guidance. Define the format, tone, when to use tools, and what to do if something fails. - Give it memory AI models forget — fast. Add memory so your agent remembers what happened in past conversations, knows user preferences, and keeps improving. - Connect it to real tools Want your agent to actually do something? Plug it into tools like CRMs, databases, or email. Otherwise, it’s just chat. - Assign one clear job Vague tasks like “be helpful” lead to messy results. Clear tasks like “summarize user feedback and suggest improvements” lead to real impact. - Use agent teams Sometimes, one agent isn’t enough. Use multiple agents with different roles — one gathers info, another interprets it, another delivers output. - Monitor and improve Watch how your agent performs, gather feedback, and tweak as needed. This is how you go from a working demo to something production-ready. - Test and version everything Just like software, agents evolve. Track what works, test different versions, and always have a backup plan. - Deploy and scale smartly From APIs to autoscaling — once your agent works, make sure it can scale without breaking. Why this matters: The AI agent space is moving fast. Companies are using them to improve support, sales, internal workflows, and much more. If you work in tech, data, product, or operations — learning how to build and use agents is quickly becoming a must-have skill. This roadmap is a great place to start or to benchmark your current approach. What step are you on right now?

“Your GTM Isn’t a Product—It’s a Platform.” a $21M CEO asked me: “how did Snowflake grow from zero to $2B+ in revenue in one of the most crowded categories?” my response? “they didn’t just build a product. they built a GTM system that scaled with every stage of growth.” most companies stall after finding early traction— 📌 they scale revenue, but not operations 📌 they hit product-market fit but don’t evolve 📌 they rely on one channel, one persona, or one hero rep but the best companies don’t just grow. they transform—from product to platform. and they do it with a go-to-market system. when GTM is a system, it evolves across stages: problem → product → platform so how did Snowflake do it? 1️⃣ predictable demand generation → how do we create pipeline at every stage of growth? 🟠 at problem-market fit: ✅ messaging focused on separation of storage & compute ✅ technical founders led early education + sales ✅ first customers were data engineers & architects 🟡 at product-market fit: ✅ launched an enterprise sales engine ✅ paid + partner channels activated ✅ early wins in finance and healthcare verticals 🟢 at platform-market fit: ✅ category creation: “The Data Cloud” ✅ multi-cloud strategy + marketplace fueled demand ✅ C-level, IT, and data teams engaged in the same ecosystem 🚀 Snowflake didn’t chase channels. they aligned GTM with product maturity. 2️⃣ seamless pipeline conversion → how do we turn interest into enterprise deals? ✅ sales process aligned to data transformation roadmap ✅ layered in vertical use cases + security/compliance value ✅ sales + SE + customer success teams worked in pods ✅ weekly forecast + usage reviews to spot and accelerate deals 🚀 every pipeline stage mapped to buyer readiness, not internal quotas. 3️⃣ revenue retention & expansion → how do we grow customer value over time? ✅ usage-based pricing → aligned value to cost ✅ net revenue retention (NRR) > 130% ✅ platform expansion: analytics → governance → apps ✅ integrations + marketplace drove stickiness 🚀 they didn’t just retain customers—they expanded into entire ecosystems. final thoughts 📌 if your GTM strategy doesn’t evolve with your product—you’ll stall. 📌 if you treat GTM as a one-time play—you’ll never become a platform. 📌 if you don’t invest in the system behind the growth—your wins won’t scale. Snowflake didn’t win because of one product. they won because their GTM system evolved at every stage. so i’ll ask you: 👉 is your GTM built to evolve—or are you still selling like it’s day one? let’s discuss 👇 — love, sangram p.s. follow Sangram Vajre to learn how to scale your GTM from product to platform with GTM O.S. #gotomarket #gtm #growth #b2b #sales #marketing #snowflake #platform #nrr #categorycreation

In 2025, deploying GenAI without architecture is like shipping code without CI/CD pipelines. Most companies rush to build AI solutions and create chaos. They deploy bots, copilots, and experiments with no tracking. No controls. No standards. Smart teams build GenAI like infrastructure. They follow a proven four-layer architecture that McKinsey recommends with enterprise clients. Layer 1: Control Portal Track every AI solution from proof of concept to production. Know who owns what. Monitor lifecycle stages. Stop shadow AI before it creates compliance nightmares. Layer 2: Solution Automation Build CI/CD pipelines for AI deployments. Add stage gates for ethics reviews, cost controls, and performance benchmarks. Automate testing before solutions reach users. Layer 3: Shared AI Services Create reusable prompt libraries. Build feedback loops that improve model performance. Maintain LLM audit trails. Deploy hallucination detection that actually works. Layer 4: Governance Framework Skip the policy documents. Build real controls for security, privacy, and cost management. Automate compliance checks. Make governance invisible to developers but bulletproof for auditors. This architecture connects to your existing systems. It works with OpenAI and your internal models. It plugs into Salesforce, Workday and both structured and unstructured data sources. The result? AI that scales without breaking. Solutions that pass compliance reviews. Costs that stay predictable as you grow. Which layer is your biggest gap right now: control, automation, services, or governance?

Your GTM strategy can make or break your startup's growth. The challenge? What works changes dramatically between the 3 major growth phases. After seeing hundreds of B2B startups scale (or fail), here's what I think each phase really looks like: PHASE 1: SELL TO LEARN Goal: Find the right offering and audience to focus on. (aka product/market fit) → Audience: Test multiple segments and use cases → Offering: Experiment with different features, pricing, and packaging → Messaging: Test different broad and specific use-case messaging → Channels: High-touch, low-cost (direct outreach, referrals, networking) → Operations: Stay lean, embrace the chaos of learning The truth? Phase 1 feels scattered. That's normal. You're gathering real data about what works. But eventually, you are ready to move to phase 2. PHASE 2: WIN YOUR NICHE Goal: Build a repeatable system to acquire customers in your chosen niche. (1 offering + 1 audience) → Audience: Focus entirely on one segment where you're winning → Offering: Perfect your solution for the core job to be done → Messaging: Hyper-focused on your chosen segment's challenges → Channels: Build a repeatable acquisition system with a few proven channels → Operations: Keep the team laser-focused on mastering one niche Success here means saying no to everything else until you've dominated this audience segment. PHASE 3: EXPAND TO GROW Goal: Build a new repeatable system to acquire a new audience or sell a new offering. → Audience: Add adjacent segments (horizontal growth) → Offering: OR add complementary solutions (vertical growth) → Messaging: Distinct positioning for each market/offering → Channels: Build new acquisition systems for each expansion opp → Operations: Dedicated GTM focus for each audience/offering The key? Choose either horizontal OR vertical expansion. Not both. And only add one new audience or offering at a time. --- The question to ask isn't "What's the perfect GTM strategy?" The real question is, "What's the right GTM strategy for my phase?" Look at your company today: → Are you still searching for what works? → Have you found it but haven't mastered it? → Or are you ready to expand what's already working? Your answer tells you exactly where to focus next. Save this framework for your next strategy session. --- P.S. This framework was inspired by a post by Robert Kaminski 🎯 on startup GTM phases a few months ago. His insights helped me refine my thinking on this. #startupgrowth #gtmstrategy

Too many teams accept data chaos as normal. But we’ve seen companies like Autodesk, Nasdaq, Porto, and North take a different path - eliminating silos, reducing wasted effort, and unlocking real business value. Here’s the playbook they’ve used to break down silos and build a scalable data strategy: 1️⃣ Empower domain teams - but with a strong foundation. A central data group ensures governance while teams take ownership of their data. 2️⃣ Create a clear governance structure. When ownership, documentation, and accountability are defined, teams stop duplicating work. 3️⃣ Standardize data practices. Naming conventions, documentation, and validation eliminate confusion and prevent teams from second-guessing reports. 4️⃣ Build a unified discovery layer. A single “Google for your data” ensures teams can find, understand, and use the right datasets instantly. 5️⃣ Automate governance. Policies aren’t just guidelines - they’re enforced in real-time, reducing manual effort and ensuring compliance at scale. 6️⃣ Integrate tools and workflows. When governance, discovery, and collaboration work together, data flows instead of getting stuck in silos. We’ve seen this shift transform how teams work with data - eliminating friction, increasing trust, and making data truly operational. So if your team still spends more time searching for data than analyzing it, what’s stopping you from changing that?