Insights

Field notes from production AI.

Written by the engineers who shipped the systems. No vendor pitches.

ARCHITECTURE · 12 MIN · MAY 2026

RAG that doesn't hallucinate: precision-first retrieval

Why recall-tuning your retriever is the wrong default, and how to instrument for it. With code.

EVALUATION · 9 MIN · APRIL 2026

Golden sets are the only AI roadmap that matters

How to bootstrap an evaluation discipline before you've shipped your first feature.

AGENTS · 14 MIN · MARCH 2026

The four agent failure modes (and how to gate them)

Runaway tool-use, context collapse, prompt-injection, hallucinated tool-arguments. What an enterprise agent must defend.

GOVERNANCE · 11 MIN · MARCH 2026

Aligning an AI programme with NIST AI RMF without slowing engineering

The minimum viable controls — and what to defer until you have a use-case actually running.

MLOPS · 10 MIN · FEB 2026

Prompt versioning, the way you'd version code

A pragmatic versioning + rollout model that gives you the regression-gate discipline you already have for code.

CASE NOTES · 16 MIN · FEB 2026

What changed when we added a reranker

An evaluated, before/after writeup of a Cohere reranker added to a production KYC retrieval system.

ARCHITECTURE · 13 MIN · JAN 2026

Why we keep choosing pgvector first

An honest comparison: pgvector, Qdrant, Pinecone — and when each one actually pays off.

AGENTS · 8 MIN · JAN 2026

Don't fine-tune until you've earned it

A flowchart that has saved more than one client a six-figure mistake. With caveats.

EVALUATION · 15 MIN · DEC 2025

LLM-as-judge: when it works, when it lies

Calibration data from 11 production deployments. Where LLM-judge is reliable, where it fails open.

One deep article per week.

Subscribe for engineering-only content from the teams that ship enterprise AI.