Notes from the lab.
Essays on building AI systems, shipping product, and the infrastructure behind them, from the Kodex team.
How to vet a development agency
Polished answers are cheap. When you vet a development agency, the useful signal is not whether they have answers. It is whether their answers have edges.
AI agents in production: the parts nobody demos
Every demo works on a clean input. Production runs on mess — missing context, adversarial users, and the long tail of edge cases. This is what it actually takes to put an agent on real traffic and keep it there.
RAG that actually answers: retrieval is a product problem
Your model isn't the bottleneck. Your retrieval is. Most RAG systems fail because the data layer was designed for search, not for answers. Here's the pipeline that holds up.
LLM security is not a prompt problem
Prompt injection gets the headlines. It's also the least of it. The real risk in an LLM system is agency — the tools you gave the model — and the data it can reach. Here's the threat model we actually use.
Platform engineering that earns its keep
Internal platforms fail when they're built for the engineering team instead of for the business. The ones that earn their budget solve one problem: how fast a product engineer goes from day one to 'it's in production'.
Shipping software without the compromise
'We had to cut corners to ship' is a story we hear constantly. It's also usually false — the corners got cut long before the deadline, in the scope, in the reviews, in the definition of done. Here's the delivery discipline that avoids it.
Observing LLM systems: beyond tokens and cost
Your error rate says nothing about whether your LLM product is getting better. Quality in an LLM system is a moving target, and most dashboards can't see it. Here's what to log, trace, and alert on.