Things we've learned, written down.
Field notes from real engagements — production gotchas, AI in the wild, methodology.
Four habits that turn an agency engagement into a product win
Most agency engagements fail not because the code is bad but because of the four shapes the engagement takes around the code. Here's what we keep doing that turns this work into the kind clients re-hire us for.
Shipping a RAG system that survives production traffic
A working demo and a production RAG system are two completely different artifacts. Here are the eight things we now always do, learned the hard way across half a dozen engagements.
11 min read
What AI can vibecode — and what it can't ship to production
AI is excellent at producing prototypes that look like working software. It is not yet good at producing software that survives a Monday morning of real users. Here's the gap, and how we close it.
9 min read
Migrations you can actually roll back
Nearly every team will tell you their migrations are reversible. Try it on a Tuesday afternoon. Here are the patterns that make rollback a routine event instead of a Saturday emergency.
11 min read
Cost ceilings for LLM features are an engineering problem, not a finance one
Most teams treat the monthly LLM bill as something for finance to escalate. It is an engineering problem with engineering levers, and the time to design for it is day one — not month three.
10 min read
The observability minimum that earns its keep
Enterprise vendors oversell observability and small teams underbuild it. There's a two-week setup that pays back forever, and a lot of stuff past that which only earns its keep at scale.
8 min read
Stack-agnostic does not mean no defaults
The most common misreading of "we pick the stack per engagement" is that we have no opinion about anything. We do. The opinions are just the floor, not the ceiling.
8 min read
Schemas that survive the second feature
The most consistent failure mode of AI-built systems is the database schema. It looks reasonable for three use cases and becomes unworkable the moment the fourth arrives.
9 min read
Authentication that survives an OWASP pass
An AI model will scaffold a login form in 90 seconds and it will look right. What it will not do, even with a reputable library, is survive a serious security review. Here's the list.
10 min read
The boring-infrastructure checklist we ship every project with
Every studio claims to build production-grade software. Production-grade is a checklist with sharp edges, not a vibe. Here's the one-page list we ship with every engagement.
9 min read
Take it from prototype to production.
Reply within one business day. Vibecoded MVP, AI-built draft, half-finished project, or a working product that's starting to crack — all welcome.
Start a project