Writing
Field notes on shipping AI systems that hold up in production — agent architecture and guardrails, retrieval quality, multi-tenant scaling, and how small teams close the gap between a demo and a deployment.
You Cannot Prompt Your Way to Safety
Web3 learned that you cannot secure an irreversible action with a warning message, because there is no undo. AI agents are acquiring the same property, and the industry is rediscovering the same four controls under new names.
Your First Engineering Hire Should Think Like a Forward Deployed Engineer
Forward deployed engineer job postings grew eightfold in a year. The role gets framed as an enterprise problem, but the mindset matters more at five people than at five hundred.
Failover That Loses State Isn't Failover
Two LLM providers at 99.53% each give you 99.99% combined. That arithmetic is real and it is also the easy half of the problem — because the conversation does not survive the switch.
What Breaks When Your AI Product Meets Its First Enterprise Customer
It is not the model. It is that a request stopped being a unit of cost, one customer can now exhaust another customer capacity, and someone is going to send you a security questionnaire.
Your RAG System Thinks It's Always Today
Embedding search has no opinion about time. It will hand you a policy that was superseded eighteen months ago with exactly the same confidence as the one that replaced it — and your users will not be able to tell.