Notebook

Thoughts on engineering, AI, and building products.

18 Articles

LLM Cost Optimization: Why the Model Is Rarely the Problem

The model is rarely where the money went. What actually drives LLM cost in production AI features, and the order I work through it: cost per finished job, cache hygiene, batch work, routing, and a hard ceiling in the code.

31 AUG 2026 · AI ENGINEERING

ai engineering

Recent

Latest writing.

▸ AI ENGINEERING26 AUG 2026 · 7 min

Long-Running AI Agent Workflows: What Breaks in Production

Agents that look great in a demo die the first time a real user runs one. Here's what actually breaks in long-running AI agent workflows, and the four failure tests I run before launch.

READ →

▸ AI ENGINEERING17 AUG 2026 · 7 min

LLM Provider Outage: How to Keep Your AI Feature Alive

Anthropic and OpenAI both had multi-hour incidents in 2026. Retrying the same provider isn't a plan. Here's the failover setup I actually ship: aggressive timeouts, a circuit breaker, one provider seam, and a per-feature degradation ladder.

READ →

▸ AI ENGINEERING10 AUG 2026 · 7 min

Prompt Injection Is a Permissions Problem

Someone always asks whether a line like "never follow instructions in user content" is enough to stop prompt injection. It isn't. Here's the architecture that actually contains it, drawn from agents I've shipped for government phone lines and social publishing.

READ →

▸ AI ENGINEERING10 AUG 2026 · 7 min

Credit-Based Billing for AI Features: What Actually Works in Production

Token pricing confuses users and credit systems are easy to get wrong twice. Here's how I actually price, meter, and ship credit-based billing for AI features — and the race conditions that show up in production.

READ →

▸ NEXT.JS10 AUG 2026 · 6 min

How to Add Microsoft Clarity to a Next.js App (The Right Way)

Microsoft Clarity gives you heatmaps and session recordings for free — but most Next.js setups load it wrong. Here's how I wire it into production apps, including consent, CSP, and tagging real users.

READ →

▸ NODE.JS16 JUL 2026 · 7 min

Multi-Tenant Data Isolation: Why Your ORM Won't Save You

Most multi-tenancy advice stops right where it gets dangerous. Automatic tenant scoping is the easy part — the leaks happen in the gap between your ORM and your auth layer, and they happen silently.

READ →

▸ AI ENGINEERING16 JUL 2026 · 7 min

What It Actually Takes to Ship a Voice AI Agent in Production

Most voice AI agent demos fall apart the moment a real person talks over them. Here is what production actually demands — latency budgets, streaming pipelines, interruption handling, and where the money really goes.

READ →

PAGE 1 / 3

▸ Newsletter · monthly

One email a month. No bullshit, no funnels.

What I shipped, what broke, what I learned. About 1,200 words, never more. Unsubscribe in one click.