Posts

The promise of multimodal AI was always that you could throw anything at a model - an image, a voice recording, a video clip, a code screenshot - and get useful output back. In 2026, that promise is largely delivered, but the details matter enormously depending on which model you pick and what you are actually trying to do. This is a practical guide to building with multimodal models today, with real numbers on latency, cost, and accuracy. ...

OpenTelemetry unified traces, metrics, and logs under a single vendor-neutral standard. In 2026, it is the default choice for observability, and proprietary agents are becoming optional.

In 2023, “prompt engineer” was a real job title. People built careers around knowing that “think step by step” improved reasoning, that XML tags helped Claude, and that role-playing made GPT-4 more creative. By 2026, all of that is either baked into the models, handled by frameworks, or irrelevant. What replaced it is more powerful and more durable: programming with LLMs instead of whispering to them. Why Manual Prompt Crafting Stopped Scaling The fundamental problem with prompt engineering was always that it was artisanal. You write a prompt, test it on 10 examples, tweak a word, test again. It is the software equivalent of hand-tuning a carburetor - it works, but only until something changes. ...

Retrieval Augmented Generation has gone from a research concept to the default architecture for building LLM applications. But somewhere along the way, an entire ecosystem of snake oil grew around it. Let me separate what actually works from what is just vendor marketing. Naive RAG Is a Solved Problem - and It Is Not Enough The basic RAG pipeline - chunk documents, embed them, retrieve top-k, stuff into context - works fine for simple Q&A over a small corpus. If you have fewer than 10,000 documents and your queries are straightforward, naive RAG with any decent embedding model will get you 80% of the way there. ...

Running LLMs on your own hardware has gone from a novelty to a legitimate production strategy. Ollama turned what used to require a PhD in CUDA optimization into a single command. But knowing which model to run, on what hardware, and with what quantization is the difference between a usable local LLM and a frustrating toy. Here is the complete guide for 2026. Why Local LLMs Matter in 2026 The case for local inference has only gotten stronger: ...

There is a pattern that has become endemic among developers who use AI tools: paste a problem into ChatGPT, copy the output, run it, fix the errors by pasting them back, repeat until it works. The code ships. The developer learned nothing. The next similar problem takes just as long. This is not AI-assisted development. It is outsourced typing with extra steps. Why Copy-Paste Creates Fragile Code Code that a developer does not understand is code that cannot be maintained. When the AI generates a solution using a pattern the developer has not internalized, three things happen: ...

Every AI startup in 2025 shipped an “agent” demo. Most of those agents broke in production within the first week. The gap between a compelling demo and a reliable agent system is enormous, and the framework you choose determines how much of that gap you have to bridge yourself. After building agent systems that handle real workloads - not chat demos, not toy examples, but systems that run unsupervised and process thousands of tasks per day - here is what I have learned about the major frameworks and the patterns that actually work. ...

The gap between engineers who use AI effectively and those who do not is no longer about prompt cleverness. It is about workflow structure. Senior engineers in 2026 are not typing better prompts - they are running a repeatable process that compounds output quality across an entire session. Here is exactly how that process works. Step 0: Context Setup Before Anything Else The single highest-leverage action is setting up project context before writing a single prompt. This means three things: ...

Claude Code reads a file called CLAUDE.md before every interaction. This file is the difference between an AI that writes code you immediately delete and one that writes code that fits your codebase. Most developers either skip it entirely or stuff it with so much text that Claude ignores half of it. Both approaches waste money and produce bad output. What CLAUDE.md Actually Does When Claude Code starts a session, it loads CLAUDE.md from the project root into its system context. Every token in that file counts against the context window and costs money. The file is advisory - Claude will follow it roughly 80% of the time, not 100%. This distinction matters and is covered more in the hooks post. ...

WebRTC promises peer-to-peer video, but the reality involves STUN servers, TURN relays, and architecture decisions that determine whether your calls actually connect. Here is how it all works.

Every AI coding session has a hidden resource that most developers ignore until it breaks: the context window. It is the total number of tokens the model can hold in memory at once - the prompt, the conversation history, every file read, every tool result, every response. When it fills up, bad things happen. Not dramatic failures. Subtle ones. The kind that waste hours. Understanding context as a finite resource - and managing it like one - is the single biggest efficiency gain available in AI-assisted development today. ...

MCP (Model Context Protocol) servers give Claude Code the ability to interact with external systems - browsers, databases, APIs, cloud services. Instead of pasting query results into chat or describing what a webpage looks like, the right MCP server lets Claude see and act on these things directly. But each server adds tool schemas to the context window, and not every server is worth the cost. Here is what is actually useful, how to set each one up, and when to skip MCP entirely in favor of CLI tools. ...

Remember learning to drive? The first week, every mirror check was a conscious decision. Turning at an intersection meant mentally running through a checklist - indicator, brake, check mirrors, turn wheel, accelerate. An hour behind the wheel left you mentally cooked. Now you drive for hours while having a conversation, planning dinner, or lost in a podcast. The same task. The same brain. Completely different experience. That shift is not metaphorical. It is a physical rewiring that happens inside the brain for every skill ever learned. And understanding the mechanism changes how you approach learning anything. ...

Every team building with LLMs eventually hits the same wall: the demo costs $0.02 per request, but production costs $0.50. The gap between prototype and production pricing is where most AI budgets die. Here is a transparent breakdown of what LLMs actually cost to run in production in 2026, and how to cut those costs without sacrificing quality. Token Pricing - The Raw Numbers As of March 2026, here is what the major providers charge per million tokens: ...

The vector database market exploded in 2023, consolidated through 2024-2025, and has now settled into a clear hierarchy. If you are building a production vector search system in 2026, your real choices come down to three: Pinecone for managed simplicity, Weaviate for flexibility, and pgvector for teams that refuse to add another database to their stack. Here is how they actually compare when you push past the marketing. Architecture - Fundamentally Different Approaches Pinecone is a fully managed, purpose-built vector database. You never see a server. You get an API endpoint, you send vectors, you query vectors. Under the hood, it runs a custom distributed architecture optimized exclusively for approximate nearest neighbor (ANN) search. Since their 2025 serverless rewrite, Pinecone separates storage and compute aggressively - you pay for what you query, not what you store. ...

AI code generation in 2026 is genuinely impressive. Models write working implementations, catch bugs, refactor with precision, and produce boilerplate at machine speed. The temptation is to route everything through AI. That temptation leads to a specific class of failures - not failures of code correctness, but failures of judgment. The pattern is consistent: AI produces code that is locally correct but globally wrong. Understanding where this breaks down is the difference between using AI effectively and letting it make decisions it is not equipped to make. ...

Event-driven architecture is not a buzzword anymore. It is how teams decouple services, handle eventual consistency, and build systems that scale without tight coordination between teams.

The conventional wisdom is simple: SQLite is for development and mobile, Postgres is for production. This was true for a long time. It is becoming less true every year. SQLite processes more queries per day than all other database engines combined. It ships in every phone, every browser, every Mac, and most Linux distributions. But the interesting shift is not about its ubiquity - it is about a growing ecosystem of tools that solve the three problems that kept SQLite out of production servers: replication, backups, and multi-node access. ...

You just shipped a new service. It works. The demo went great. Leadership is happy. Then the bill arrives - not just the AWS invoice, but the full picture. The on-call rotation that burned out two engineers. The three-day debugging session caused by a misconfigured load balancer. The database that needed emergency sharding at 2 AM because nobody did the capacity math upfront. Your cloud bill is the most visible cost of running production systems. It is also the smallest one. ...

You’re three months into a new project. You have a PostgreSQL database for your core data, Redis for caching, Elasticsearch for search, MongoDB for “flexible” documents, and TimescaleDB for metrics. Five different systems, five different failure modes, five different backup strategies. Your on-call rotation is a nightmare. Here’s the thing - Postgres alone could have handled four of those five jobs. Maybe all five. I’m not saying Postgres is the only database you’ll ever need. I’m saying it should be the last database you add to your stack, not the first one you try to replace. There’s a reason it keeps winning. ...