Insights
Essays on how to tell whether an AI system works: what to measure, what to test, and what to ask before it reaches production.
Spark Declarative Pipelines: What You Own After You Adopt Them
An evaluation of Apache Spark 4.2.0 Declarative Pipelines, measured on a single-node OCI lab, checked against vendor documentation, with the architecture consequences for teams already running Airflow and dbt.
The Use-Case Lottery
Enterprise AI portfolios fail at selection, not execution. Most are stacks of lottery tickets bought by whoever pitched loudest, and the fix is portfolio governance, not more pilots.
From Projection to Proof: A Framework for Measuring AI Productivity Gains
Calling the productivity measurement problem a Luddite argument misses the point. AI can deliver real gains. Most organisations are measuring for the wrong trajectory and getting false negatives as a result. Here is what rigorous looks like.
Nobody Bought Productivity
Billions are being spent on AI tools and nobody can measure the productivity gains. The measurement debate has been missing the point. The actual investment thesis, named honestly, is something the budget memo cannot say.
Tools as the Only Interface: How to Make AI Agents Observable by Design
Every agent action is a tool call. Nothing happens outside of tools. This design choice made 15 days of autonomous behavior fully auditable. Enterprise agentic systems need the same principle.
The Agency Maturity Map: Why Most Enterprise AI Is Operating in the Red Zone
There are four stages of enterprise AI maturity and three gates between them. Most organizations skip the gates. That is the only thing you need to know to explain why most enterprise AI initiatives stall, drift, or quietly fail.
The Compliance Trap: Why Regulated Industries Keep Getting AI Wrong
Healthcare, finance, and insurance organizations are overindexing on two things that feel like solutions but aren't: domain expertise and vendor platforms. Here is what determines whether enterprise AI works in regulated environments.
The Token Trap: Why Your AI Adoption Metrics Are Measuring the Wrong Thing
Counting tokens to measure AI adoption is like measuring software quality by lines of code. Counting outcomes is too slow and too late. Here is what works.
What No Benchmark Can Measure: 15 Days of Autonomous AI Agents in the Wild
Benchmarks test isolated capabilities. Season 1 of Emergence World ran five AI societies for 15 days each, and what collapsed and what survived tells you more about production AI than any leaderboard.