Blog
News, guides, and updates from Curtis Foster.
AI agent observability: the signals that reveal what your system is really doing
AI agents rarely fail in one obvious place. Learn which traces, metrics, and signals expose the real failure before your users do.
The first week after launching an AI product: what broke, what surprised me, and what I fixed
The first week after launching an AI product exposes the gap between a working demo and a dependable service. Here are the failures I found and the fixes that mattered.
AI product requirements: how to write a useful PRD for an uncertain AI feature
AI features need PRDs that describe uncertainty, evidence, and failure paths. Here is a practical way to write one your team can actually build.
7 questions to ask before giving an AI workflow access to customer data
Before an AI workflow reads customer records, ask seven questions about access, retention, vendor use, errors, and accountability.
What building Ziggyloo taught me about designing an AI tutor product for real learners
Building Ziggyloo taught me that an AI tutor is not a chatbot with homework mode. It is a learning system that must manage attention, uncertainty, and progress.
The AI feature that looked smart until a real user asked a follow-up question
A polished first answer can hide a broken conversation. Here is how to design, test, and ship AI features that survive real follow-up questions.
How I'd hire an applied-AI engineer in 2026
The best applied-AI engineers do more than call a model API. Here is how I would test judgment, engineering depth, and product sense before making an offer.
Per-call cost telemetry: how to stop AI bills from surprising you
AI bills become manageable when every model call has an owner, purpose, and recorded cost. Here is how to find waste before finance finds it.
Multi-agent systems: when they help, and when one good agent wins
Multi-agent systems can divide complex work, but they also multiply latency, cost, and failure points. Here is how to decide when one good agent is enough.
What a solo founder learned shipping AI from 0 to 1
Building Ziggyloo and Zentoko alone changed how I think about scope, reliability, cost, and the work that turns an AI demo into a product.
AI memory architecture: the five layers behind reliable LLM products
AI memory is not one database or a longer context window. It is five separate layers, each with a different job, cost, and failure mode.
The eval loop is the product: measuring whether your AI is actually right
A production AI system is only as good as its measurement loop. Learn how to build AI evals that catch bad answers before users do.
Subscribe to Curtis Foster
Get new posts in your inbox. No spam, unsubscribe anytime.