AI

🎯 Intro Point an agent at a catalog with 400 tables and ask it a simple question. It reads schemas. Then samples rows. Then guesses at a join. Fourteen tool calls later it hands you a number that’s off by 12%, and you have no idea which step went wrong. Now hand the same agent three sentences: what a customer is, which table is the trusted one, and how the company counts “active.” One query. Correct answer. ...

🤔 The enterprise-wide enigma Ask three people on your data team what “total revenue” means and you’ll get three queries. One filters out returns. One rounds before summing. One groups by order date, another by ship date. All three are technically correct, and none of them match. This isn’t a data quality problem. The data is fine. The problem is that every dashboard, notebook, and BI tool re-implements the aggregation logic from scratch, and nothing stops those implementations from drifting apart from each other. ...

Your agent makes a tool call. The tool returns 40 rows of JSON. That JSON gets stuffed back into the model’s context, the model thinks for a bit, calls another tool, gets another 40 rows back, and on it goes. By iteration 10, your prompt is mostly curly braces, quoted keys, and the same field names repeated hundreds of times. You’re paying for every single one of those tokens. And you’re paying for them on every subsequent model call in the run, because tool results stick around in the conversation history. ...

🎯 Introduction “I’ll just throw it in DynamoDB.” I’ve heard this line dozens of times from engineers building AI agents. It sounds reasonable. You need to store conversation history, maybe some user preferences. DynamoDB is fast, scalable, and you already use it. How hard could it be? Here’s the thing: agent memory isn’t just storage. It’s what separates an agent that forgets everything you said five minutes ago from one that actually remembers your last conversation, pulls up context from three weeks ago, and knows which details matter and what data is fresh. ...

Remember when chatbots were the future of customer service? Fast-forward five years, and we’re already talking about AI agents that can handle your entire sales pipeline, analyze market data, and even make procurement decisions without human intervention. This isn’t science fiction. Companies like Salesforce and Microsoft are rolling out agentic AI systems that go far beyond answering “What are your hours?” They’re building virtual employees that think, plan, and execute complex business tasks. ...

What if I told you that you could run vector searches directly on S3 without spinning up a single database or compute cluster? For years, we’ve been stuck with a painful pipeline: extract data from S3, chunk it, generate embeddings, load everything into OpenSearch or Pinecone, and manage all that infrastructure. Amazon just changed the game with S3 Vectors – it’s S3 that can do vector math natively, no compute engine required. This means up to 90% cost savings and zero infrastructure management. Let me show you exactly how this works and why it might replace your vector database entirely. ...

Introduction Have you ever wondered how Pinterest finds visually similar images or how Google Photos recognizes faces across thousands of pictures? The technology that powers these features isn’t magic—it’s vector similarity search. Today, modern vector databases make it possible for developers to build these powerful visual search capabilities without needing a PhD in computer vision. In this post, I’ll guide you through the process of building your own image similarity search engine. We’ll cover everything from understanding vector embeddings to implementing a working solution that can find visually similar images in milliseconds. ...