Architecture

Every Architecture article we've published — field notes on building and shipping production AI systems, written by the engineers who build them.

8 articles

Patrick Waweru10 min read

When ten thousand people ask the same question at once

A cache answers questions it has seen before. During an incident it has seen none of them, and every request misses at the same moment. How to do the work once instead of four thousand times.

Oct 8, 2026 Read article

George Onyango10 min read

Cost, speed, quality. You get two.

Every AI system trades one of the three against the other two, and the expensive mistake is refusing to choose. How to decide which corner you are giving up - starting from who is waiting for the answer.

Oct 7, 2026 Read article

Harrison Itotia10 min read

Ordinary systems fail loudly. These fail plausibly.

A language model system returns 200, the answer is fluent, every dashboard is green, and it is wrong. The four-column table that forces you to admit which failures you have no control for.

Oct 6, 2026 Read article

George Onyango10 min read

The human isn't a failure of automation

Three places a person belongs in an AI system, how to design the job so it still works in six months, and the incident where every automated check passed and a person caught it anyway.

Sep 30, 2026 Read article

Patrick Waweru11 min read

Fast AI products do the work before you ask

You can't make a language model think faster. You can stop it thinking while somebody waits. The precomputation trick behind every AI product that feels instant, in plain English.

Sep 29, 2026 Read article

Moses Otieno12 min read

Microservices that survive contact with production

Nine microservices best practices that separate a fleet you can hand to an auditor from a distributed monolith: contract-first APIs, one trust boundary, etags over last-write-wins, and bounds on everything.

Sep 4, 2026 Read article

George Onyango11 min read

Why 95% of agent pilots die — and the boring engineering that ships the other 5%

The failure rate is real, and it is almost never the model. Five unglamorous engineering decisions separate the agents that reach production from the ones that die in the demo.

Aug 19, 2026 Read article

Harrison Itotia10 min read

Why narrow specialists beat one 30-tool agent

The fastest way to stall an agent in the pilot is to hand one model thirty tools. Decision paralysis is an architecture problem — and a small team of narrow specialists is the fix that ships.

Aug 12, 2026 Read article