Cloud native for AI workloads: what’s actually different
Your Kubernetes cluster was designed for stateless HTTP handlers. Your AI workload is none of those things.
Engineering
Engineering
Your CI/CD pipeline can’t test a model. Here’s what to build instead.
The pipeline that deploys your web app runs 400 unit tests. The pipeline that deploys your model runs… what, exactly?
Engineering
The 8 Fallacies are 30 years old. They’re worse with LLMs.
Peter Deutsch wrote the fallacies in 1994. LLM-backed architectures violate all eight, sometimes in the same request.
Engineering
You probably want a modular monolith
We build microservices. We also talk clients out of them fairly often, and this is the argument we use.
Engineering
The 7 Rs are fine. The problem is everyone picks rehost.
An assessment that returns the same answer for every application is not an assessment.
Product
Why every drug in Auscura carries its citation
Building the prescription module meant deciding what "evidence-based" is allowed to mean.
Engineering
What 30-year-old COBOL taught us about writing agent specifications
Extracting business rules from a legacy policy administration system turns out to be the same problem as specifying an agent.
Research
The EU AI Act’s high-risk deadline moved. Your timeline didn’t.
The Digital Omnibus pushed Annex III high-risk obligations to December 2027. That extension is smaller than it looks.
Engineering
Build the evaluation harness before you build the agent
Every agent programme we’ve seen fail shared one property: nobody could say whether the agent was getting better.
Product
Stayligent: why we built a PMS that acts instead of records
Hospitality software has spent fifteen years adding integrations to a system of record. We think the property needs an operating system.