Agentic Systems
The Hidden Loop Behind Every AI Answer I used to think AI coding tools just had a very good search engine inside. Then I looked at what actually happens when you ask one a question.
Read note → Agentic Systems
The Model Isn't What Limits Your Agent The capability of an AI agent is not set by the model. It's set by the tools.
Read note → Agentic Systems
Understand, Plan, Execute: The Step Most People Skip Most engineers jump straight to execute mode. That's usually the expensive mistake.
Read note → Retrieval
Why Your Agent Only Reads 10 of 40,000 Files Your codebase has 40,000 files. The AI agent reads maybe 10 of them before answering your question. That's not a bug. That's the design.
Read note → Engineering Practice
AI Wrote the Code. AI Reviewed the Code. Nobody Caught the Bug. This happens more often than people admit — and it's not the AI's fault. It's how we're using it.
Read note → Production AI
Built in a Day. Broken in Production. When people see one-day AI demos, they're not really wondering if engineers will be replaced. They're asking: if anyone can build this now, what still truly requires expertise?
Read note → Retrieval
Retrieval Fails Before the Model Ever Sees Your Question Most people assume document AI fails because the model is weak. Often, the problem starts earlier — in retrieval.
Read note → Model Internals
Is "Reasoning" Even the Right Word? I kept saying 'the model reasoned through it.' I'm not sure that's the right word anymore.
Read note → Code Intelligence
Three Questions, Three Different Searches Three different questions to a coding assistant. Three fast answers. Same chat window. I assumed the same search was running each time. It probably isn't.
Read note → Evaluation
The Answer Was Right. The Evidence Wasn't Real. An AI agent gave the right answer. But the evidence it quoted wasn't real. The answer passed the 'is this correct?' test. It failed the 'did it actually use real evidence?' test.
Read note → Engineering Practice
Code Generation Got Fast. Review Confidence Didn't. We're building an entire ecosystem around generating code. Now we're building another ecosystem to review the code generated by the first one.
Read note → AI Security
SQL Injection Has a Fingerprint. That's Why WAFs Work. Every web attack for 20 years had something in common: syntax. That insight built an entire industry.
Read note → AI Security
The Attack Wasn't in the Prompt. It Was in the Document. The attack didn't come from a hacker probing your API. It came from a document your AI agent was asked to summarize.
Read note → AI Security
Your WAF Can't Read Intent. Prompt Injection Doesn't Care. Your WAF can often block SQL injection because SQL injection has recognizable patterns. Prompt injection doesn't.
Read note → Production AI
The Model Got Smarter. Our Results Got Worse. We upgraded to a newer model, clearly better than what we had before. Same code, same data, same test cases. Instead, some results got worse.
Read note → Production AI
Is Your Tool Showing the Model Reality, or Your Interpretation of It? Every AI tool I build now starts with the same question. Am I giving the model access to reality? Or access to my interpretation of reality?
Read note → Production AI
Not All AI Engineering Has the Same Shelf Life Some of it becomes more valuable every time models improve. Some of it slowly loses value — and the tricky part is that both look equally important while you're building them.
Read note → AI Governance
You Can't Govern the AI You Can't See The biggest AI governance problem isn't enforcement. It's visibility.
Read note → AI Governance
The Real Question Isn't Which AI Gateway to Buy Every AI governance discussion eventually turns into: 'Which AI gateway should we use?' I don't think that's the first question.
Read note → Context Engineering
The Goal Isn't More Context. It's the Right Context. As AI models get larger context windows, I've found myself thinking differently about context. It's becoming less about how much we can provide, and more about which context should always be available.
Read note → Retrieval
Fine-Tuning vs RAG: Why I Chose Not to Fine-Tune Building an AI assistant for RAG might seem straightforward. Creating something production-ready comes with its own set of challenges and nuances that aren't always obvious at first.
Read note → Retrieval
Ingesting PDFs, CSVs, and JSON Into One RAG Pipeline An AI assistant's ability to provide insightful answers relies heavily on the data it has access to. But the data isn't always uniform.
Read note → Retrieval
FAISS vs Chroma vs Milvus: What I Learned Hands-On Choosing the right vector database depends on your project's specific needs, including scalability, metadata handling, and deployment preferences.
Read note → Retrieval
The Overlooked Problem: Duplicate Data Ingestion Re-ingesting identical data quietly costs you storage, speed, and correctness. A simple hash check fixes most of it.
Read note → Retrieval
Four Qdrant Assumptions That Didn't Survive Real Data A few early assumptions about vector databases didn't hold up once real data came into play.
Read note → Agentic Systems
Demo Was Easy. Production MCP Server Was Not. Wrapping every API into its own tool works fine in a demo. It falls apart once a product has 100+ of them.
Read note → AI Memory
AI Memory Isn't About Remembering More. It's About Remembering the Right Things. Before something becomes persistent memory, there's an implicit question: can this information be recovered later?
Read note → AI Memory
Repositories Remember What. AI Memory Should Remember Why. A repository is excellent at explaining what exists. What often disappears over time is the reasoning behind those decisions.
Read note → AI Memory
Good AI Memory Shouldn't Learn From Every Conversation A single conversation often reflects the task you're working on today, not a long-term preference. If every interaction became permanent memory, the system would quickly become inconsistent.
Read note → AI Memory
The Hardest Thing About AI Memory Isn't Remembering. It's Forgetting. If an AI assistant forgets something useful, we can usually explain it again. If it continues reasoning from assumptions that are no longer true, the problem is much harder to notice.
Read note → AI Governance
AI Traffic Changed. Governance Disciplines Evolved. Almost every AI gateway product introduces capabilities that sound new. A simple question changed how I looked at it: what engineering problem is this actually solving?
Read note → AI Governance
New Vocabulary, Old Engineering Disciplines Every major technology shift brings a new vocabulary. Looking past the terminology often reveals a familiar engineering discipline underneath.
Read note → Model Internals
It's Not Just the Harness The prompts, tools, and workflow stayed identical. The only thing I changed was the underlying model, and the investigation behaved completely differently.
Read note → Model Internals
The Fundamentals I Memorized in College Finally Made Sense I'd studied vectors, matrices, and probability years ago. I only understood them once I started implementing an LLM's building blocks myself.
Read note → Production AI
Quality Doesn't Degrade Gradually. It Drops Off a Cliff. I assumed model quality would degrade visibly in a long session. What actually happens is different, and the threshold arrives earlier than it feels like it should.
Read note → Production AI
The AI Prototype Worked in Three Days. The Integration Took Months. The demo connected to two clean APIs. The enterprise version needed an HR system, a CRM, a finance platform built in the early 2000s, and a data warehouse that only refreshes overnight.
Read note → Production AI
The Paper Trail: Why Observability Is a Deployment Prerequisite When you're the only user of an AI tool, you don't need to explain why it did what it did. Enterprise deployment removes that luxury entirely.
Read note → Retrieval
How Vectorless RAG Actually Navigates a Document Most retrieval systems compress your question into a point in space and find the nearest match. Vectorless RAG doesn't do that. It navigates.
Read note → Retrieval
Vector RAG and Vectorless RAG Aren't Competing for the Same Job The mistake I kept seeing was reaching for the fast retrieval approach when the accurate one was actually needed.
Read note → Model Internals
Why the Model Doesn't Miss the Quieter Concepts in Your Question Most questions have more than one concept, and the concepts don't carry equal weight. Here's the mechanical reason the quiet one still gets heard.
Read note → Model Internals
What Q, K, V Actually Mean Inside every transformer, every token plays three roles at once. Understanding what those roles are makes the model's behavior much less mysterious.
Read note → Code Intelligence
Code Isn't Flat. That's the Whole Problem. Most retrieval systems were built for prose, where structure is mostly decorative. Code is the opposite, and chunking it like prose breaks the thing that makes it meaningful.
Read note → Evaluation
The Answer Was Correct. Nobody Could Act on It. The semantic score was high. The faithfulness check passed. The resolution recommendation was still unusable at 2am during a live incident.
Read note → AI Security
The Gap Between a CVE and a Code Fix Used to Close in Hours Virtual patching closed a two-week exposure window down to hours, because the exploit had a syntactic fingerprint. That's exactly what's missing when the attack is a sentence.
Read note → AI Security
The Blast Radius Grows With the Permissions A chatbot that gets jailbroken is embarrassing. An agent with database access, email permissions, and a file system connection getting injected is a different category of problem.
Read note → AI Security
The Perimeter Moved. Security Hasn't Caught Up Yet. In traditional web security, the request coming in is the event you care about. In an agent architecture, the request is just the beginning.
Read note →