Blog
Written by the engineers who did the work, usually because something surprised us and we wanted it written down before we forgot.
Shipping intelligence into a revenue-carrying product is an architecture problem before it is a model problem. The sidecar pattern, feature flags, and why you want a switch that turns the whole thing off.
A golden dataset and an automated eval suite are the difference between an agent you can change and one nobody dares touch. What to measure, and how to make regressions fail CI.
Rewrites fail because they ask a business to stand still for a year. Routing capability away from a legacy system one endpoint at a time, with a rollback path at every step.
The 28% return-rate reduction came from per-SKU garment measurements and reason-coded returns, not from a bigger model. A note on where the leverage really sits in retail AI.
The agent books, reschedules, chases and bills. It drafts clinical notes and never files them. How we decided which decisions an autonomous system was allowed to own.
Polyglot stacks are defensible at scale and expensive before it. Why FastAPI, Celery and PyTorch in one repo keeps a five-person team faster than a fifteen-person one.
Full articles are being migrated across. Want one of these as a conversation instead? Get in touch
Start Your Project
Most of these posts started as a client problem. Send us yours and we will tell you which lesson applies.
Tell us what you are trying to build. A senior engineer reads every brief that comes through here.
Ithrive AIChat
Answers from our own case studies
I can answer questions about what we build, walk you through any of our ten case studies, and put you in front of an engineer when it is worth it. What are you working on?
AI assistant — it can be wrong. For anything binding, email hello@ithrivesoftware.com.
