Practical articles on building production software — from type-safe AI pipelines to legacy migrations and edge deployments.
Fine-tuned small models outperform zero-shot GPT-4 on 80% of classification tasks at 10–100x lower cost. LoRA, DPO, and fine-tuning-as-a-service changed the calculus. Here's the decision framework.
Text-to-SQL is one of the most useful AI features you can build — and the most dangerous if done carelessly. Here's the five-layer security architecture for production.
Pre-deployment testing increasingly fails to predict real-world AI behavior. Here's a four-category framework for systematically finding your AI's failure modes before your users do.
The gap between "works on my laptop" and "runs reliably 24/7" is wider than most developers realize. State persistence, error recovery, resource budgets, and the five problems nobody warns you about.
73% of teams use AI daily but productivity gains are only 10%. The gap between 3x gains and zero gains is engineering judgment — and AI is making it more valuable, not less.
A production AI agent was tricked into issuing $47,000 in unauthorized refunds. Traditional authentication fails for agents. Here's the secure architecture pattern.
The multimodal AI market hit $3.85B in 2026. Google's Gemini Embedding 2 maps text, images, and audio into one vector space. Here's how to add vision to your app with practical code.
Even calling an LLM API triggers legal obligations if you serve EU users. Fines up to €35M. The major deadline is August 2, 2026. Here's the practical checklist.
In 2022, 60% on MMLU required 540B parameters. By 2024, 3.8B parameters hit the same score — a 142x reduction. SLMs are the future of production AI, and the data proves it.
I was paying $400/month for ticket classification. A fine-tuned Mistral 7B matched the accuracy at $31/month. Here's the decision framework for when open-source beats proprietary.
Eight years of production software, distilled into one stack. React 19, TanStack Start, Cloudflare Workers, PostgreSQL, and the AI tools that save me 3 hours a day.
Your API was designed for developers who read documentation. AI agents are a different consumer — brilliant at structure, terrible at reading between the lines. Here's how to design for both.
Vector storage, conversation history, prompt versioning, usage analytics, and job scheduling — all in one database. Why I stopped reaching for six services and used PostgreSQL for everything.
A healthcare client refused to send data to external APIs. WebGPU and Transformers.js let me run a text classifier entirely in the browser — 30ms inference, zero API calls, full privacy.
Six months after launch, users said search was getting worse. The code hadn't changed. The culprit: embedding drift — model deprecation, silent updates, and corpus evolution.
My last production bug passed TypeScript strict mode and every lint rule. An LLM found it in 5 seconds. Here's how to add AI code review to your CI pipeline for $2/month.
At 2 AM on a Tuesday, the Anthropic API went down for 47 minutes. Our AI feature became a blank page. That was a $12,000 lesson in building resilient AI products.
40% of our LLM calls were near-duplicates — same question, different words, full price. Semantic caching with pgvector dropped the monthly bill from $215 to $85.
A user asked our compliance AI to write a poem about marijuana regulations. It did, beautifully. That was the wrong answer. Here are the five layers of guardrails I built afterward.
Users abandoned our AI tool — not because the output was bad, but because it took 5 seconds. Streaming, skeleton states, predictive loading, and the 500ms rule that changed everything.
Prompt engineering was never the real skill. Context engineering — designing the informational environment around the model — is what actually determines output quality. Here's what changed and why it matters.
73% of dev teams use AI coding tools daily, but org-level productivity gains are only ~10%. The reality behind the hype, the controversy, and what actually works.
Claude Code went from launch to the #1 AI coding tool in under a year. This isn't just better autocomplete — it's a fundamental shift from code completion to task completion.
From internal Anthropic experiment to Linux Foundation standard with 100M monthly downloads in one year. How Model Context Protocol became the universal connector for AI agents.
Traditional testing assumes determinism. AI agents break that assumption. Here's how evaluation-driven development, behavioral testing, and agent observability are filling the gap.
AI agent pilots doubled in 2025 but full deployment is stuck at 11%. Here's what's blocking production and a practical framework to break through.
Multi-agent AI architectures are following the same arc as microservices. Here are the design patterns, orchestration models, and protocols you need to know in 2026.
Token pricing dropped 99.7% in three years. Here's how that cost collapse is reshaping AI agent architectures and creating a new discipline of agent economics.
AI agents are learning to optimize their own code, discover new tools, and accumulate skills over time. Here's how recursive self-improvement works, why it matters, and what keeps researchers up at night.
A practical guide to building real-time AI chat interfaces in React using Server-Sent Events — from server-side stream creation to smooth client-side token rendering.
A deep dive into building LLM-powered agents that can execute real actions — from defining type-safe tools with Zod to implementing the agent loop and handling failures in production.
Why AI logic belongs on the server, how React Server Components create a natural boundary for LLM calls, and patterns for streaming AI responses to interactive client UIs.
How I built a production retrieval-augmented generation pipeline using PostgreSQL and pgvector instead of a dedicated vector database — from embedding documents to semantic search.
Why useState falls apart for AI chat interfaces, and how TanStack Store provides fine-grained reactive state management that keeps streaming UIs performant.
The adapter pattern for LLM providers — how to build a provider-agnostic AI layer in TypeScript so you can switch between OpenAI, Anthropic, and Google without touching your application code.
How to use TypeScript's type system and Zod validation to build reliable, production-grade LLM data pipelines that catch errors before they reach your users.
How to deploy AI-powered features to Cloudflare Workers — from proxying LLM API calls at the edge to using Workers AI for on-device inference, with real latency benchmarks.
How to use Zod schemas and provider-specific structured output features to get validated, type-safe data from LLMs — no more regex parsing or hoping the model follows instructions.
A practical guide to migrating legacy Blazor/.NET enterprise applications to React and TypeScript — from reverse-engineering undocumented APIs to shipping a modern frontend.
How I built a dashboard that displays AI pipeline status, processing metrics, and LLM-generated insights in real time using TanStack Query polling and optimistic updates.
Battle-tested prompt engineering patterns from production AI systems — system prompt architecture, few-shot examples, chain-of-thought reasoning, and how to version and test prompts like code.
Everything you need to know about deploying full-stack React applications to Cloudflare Workers using TanStack Start — from Vite configuration to edge-native patterns.
Lessons from building an AI-powered regulatory compliance platform from zero to MVP as a sole developer — covering RAG pipelines, prompt engineering, and payment processing.
Ask me about Kyle's skills, experience, or projects