10 Sept 2026•2 min read
Build Times Are a Culture Problem
Slow feedback loops do not just waste minutes. They change how developers work, pushing teams toward larger batches and less experimentation.
Tag
Every essay tagged Architecture, newest first. Tags are added by the writer, so this is what our authors themselves considered architecture — not an automatic match on the word.
10 Sept 2026•2 min read
Slow feedback loops do not just waste minutes. They change how developers work, pushing teams toward larger batches and less experimentation.
8 Sept 2026•5 min read
The jump from answering questions to operating software breaks most of the assumptions we built our AI features on. Here is what actually needs to change in your stack.
5 Sept 2026•4 min read
Most RAG systems fail long before the embedding model is at fault. The fixes live in chunking, freshness, and permissions, not in a better index.
3 Sept 2026•3 min read
Sending every request to your largest model is the most expensive default in modern software. A routing layer usually beats an upgrade, on both cost and latency.
2 Sept 2026•3 min read
Reversibility is the cheapest trust mechanism available to anyone shipping automation, and it is consistently the last thing teams build.
30 Aug 2026•3 min read
Models change every few weeks. Prompts get rewritten by whoever touched them last. The only durable asset in an AI product is the set of cases you can test against.
29 Aug 2026•3 min read
When work takes ninety seconds, HTTP request-response stops being an adequate model. The patterns that fix it are borrowed from job systems, not from web frameworks.
27 Aug 2026•3 min read
You cannot filter your way out of untrusted text reaching a model. What you can do is make sure that text has no authority when it arrives.
26 Aug 2026•3 min read
Traditional logs answer what happened. With probabilistic components you also need to answer why this run differed from the last one, which requires capturing inputs you probably are not storing.
22 Aug 2026•3 min read
Deploying close to users stopped being a differentiator and became the default. The interesting consequences are in how we design data, not in how fast the HTML arrives.
19 Aug 2026•3 min read
Self-hosting a capable model is now a genuine option rather than a research exercise. The decision turns on data sensitivity and load shape, not on ideology.
16 Aug 2026•3 min read
Compute demand is now large enough that power availability shapes where and when workloads run. That has started to show up in ordinary engineering decisions.
15 Aug 2026•2 min read
The moment your internal service has three consumers it has become a platform, and platforms fail differently from applications. Most teams notice too late.