Writing
Notes on building software
Honest takes on shipping products, indie hacking, and the realities of the tech industry. No fluff.
The LLM Router Pattern in 2026: Model Routing, Fallbacks, and Cost Control That Actually Works
Picking one model for your whole app is the bug. The teams shipping the best AI products in 2026 route every request to the cheapest model that can handle it, fail over when providers blink, and treat model selection as part of the app, not part of the prompt. Here is how to do it without making a mess.
Sandboxing AI-Generated Code: E2B vs Vercel Sandbox vs Modal vs Daytona in 2026
Letting an LLM write code is the easy part. Letting it run that code on a machine that touches your data is the part that should keep you up at night. Here is how the production sandboxes compare in 2026, and what actually matters when you pick one.
AI Agent Frameworks in 2026: LangGraph vs Mastra vs Vercel AI SDK vs OpenAI Agents SDK vs Pydantic AI
There are too many agent frameworks and most of the comparisons online are useless. Here is what I have actually shipped on each, where they shine, and where they will quietly cost you a weekend you did not budget for.
Generative UI in 2026: What Actually Works for Developers
Chat is a terrible interface for most things AI agents do. Generative UI is finally good enough to ship, and the patterns that work are not the ones the demos show. Here is what I have learned shipping AI features that render real components instead of walls of text.
AI Voice Agents in Production: What Actually Works in 2026
Voice agents went from "cute demo" to "real product surface" this year. Most of them still feel terrible. Here is what separates the voice AI experiences people actually use from the ones they hang up on, written from the trenches.
Generative Engine Optimization: How to Get Cited by ChatGPT, Perplexity, and Claude in 2026
Half the search traffic that used to land on blogs now lands on AI answers instead. Getting cited inside those answers is the new SEO, and the rules are different. Here is what is actually working in 2026 for indie hackers and small teams.
Problem Validation for Indie Hackers: Why You Are Building the Wrong Thing (And How to Know Before You Start)
The gap between "can build" and "should build" has never been wider. AI tools made shipping trivially fast. That is exactly why building the wrong thing faster than ever has become the new way to fail. Here is a practical validation framework for solo founders who want to work on problems worth solving.
Securing AI Agents in Production: What Nobody Tells You Before Something Breaks
A Cursor AI agent deleted a production database in nine seconds. Not because the AI was malicious, but because nobody thought carefully about what it was allowed to touch. Here is a practical security framework for running AI agents in production without handing them the keys to everything.
AI Browser Agents in 2026: Stagehand vs Browser Use vs Playwright
Most browser automation tutorials show you how to click buttons. They do not show you what happens when the button moves, the page layout changes, and the model confidently clicks the wrong thing anyway. Here is how to build browser agents that survive contact with the real web.
LLM Fine-Tuning for Developers: When RAG Is Not Enough
RAG is the answer for most retrieval problems, but there is a class of problem it cannot solve: when you need the model itself to behave differently, not just know more things. This is the practical guide to fine-tuning I wish I had before I wasted two months and four thousand dollars doing it wrong.
Multi-Agent vs Single-Agent Architecture in 2026: When the Crew Beats the Soloist
Multi-agent systems are the architecture pattern everyone is talking about in 2026 and almost nobody actually needs. After shipping both shapes in production, here is the honest framework for when a crew of agents beats a single well-prompted one, and when it just multiplies your bugs.
Structured Outputs in 2026: Function Calling, JSON Mode, and the Schema Wars
Three years ago you parsed LLM JSON with a prayer and a regex. In 2026 every major provider supports schema-constrained outputs, but they all do it differently, and the wrong choice will silently corrupt your data. Here is the field guide I wish I had before I shipped four broken integrations.
Prompt Caching in 2026: Anthropic vs OpenAI vs Gemini for Production Apps
Prompt caching is the quiet unlock that makes long context economics work in production. But every provider implements it differently, the pricing math is not obvious, and most developers are leaving 70 to 90 percent savings on the table. Here is a field guide after burning a lot of tokens to figure out what actually works.
Temporal vs Inngest vs Vercel Workflow in 2026: Picking a Durable Engine
Durable execution engines went from "interesting infra pattern" to "the only sane way to build AI agents and long-running background work" in 2026. Temporal, Inngest, and Vercel Workflow are the three I keep seeing in production. Here is how they actually compare after running real workloads on all three.
RAG vs Long Context in 2026: When to Retrieve and When to Just Stuff the Window
Claude Opus 4.7 ships with a 1 million token context window. Gemini 2.5 has 2 million. GPT-5 sits at 400k. The obvious question: do we still need RAG, or can we just paste the whole codebase into the prompt? After rebuilding two production features both ways, the answer is not what I expected.
Vector Database Comparison 2026: pgvector, Pinecone, Turbopuffer, and Qdrant
I spent the last two months running the same RAG workload across pgvector, Pinecone, Turbopuffer, and Qdrant on real production traffic. Here is what actually shipped, what broke, and which one I would pick if I were starting a new project this week.
AI Code Review Tools in 2026: CodeRabbit vs Greptile vs Vercel Agent
AI code review tools moved from novelty to mandatory in 2026. CodeRabbit is the market leader, Greptile is the technical darling, and Vercel Agent is the native pick for anyone deploying on Vercel. Here is an honest comparison after running all three against real pull requests on real codebases.
Durable AI Workflows in 2026: Why Your Next AI Feature Needs Orchestration
AI agents that talk to APIs, run for minutes, and touch external state break in ways your typical request-response code does not. Durable workflow engines like Inngest, Trigger.dev, and Vercel Workflow solve a problem most developers do not realize they have until production burns them. Here is the guide I wish I had six months ago.
AI Agent Observability: Debugging Production Agents Without Going Insane (2026)
The reason your agent works in development and quietly falls apart in production is almost never the model. It is that you cannot see what it did. Here is the observability setup that turns mystery failures into fixable bugs.
Cursor vs Windsurf vs Zed: The AI IDE Showdown (2026)
Three AI-native IDEs are competing for the seat next to your keyboard. I used each one as my daily driver for a week. Here is what actually matters when you pick between Cursor, Windsurf, and Zed in 2026.