Articles
Twice a week, on what actually breaks in production.
8 Oct 2026
Incremental refresh is cheap until the watermark drops rows. The real decision is partition grain, the rescan window, and how you prove it still reconciles.
Data engineering
8min read
6 Oct 2026
How to design an evaluation harness for agentic and LLM systems: hard asserts, graded rubrics, fixture hygiene, judge calibration and what a red run blocks.
Artificial intelligence
8min read
1 Oct 2026
Reports rot quietly: definitions drift, audiences move on, copies multiply. A concrete review habit that keeps a reporting estate honest without a rebuild.
Business intelligence
7min read
29 Sept 2026
Commodity capability, finite budget: how to decide what to buy, what to build, what to defer — and why the second year of ownership settles the argument.
Data & AI strategy
8min read
22 Sept 2026
A star schema rarely breaks. It drifts — grain, conformance, keys, history and partition boundaries all encode assumptions that quietly expire.
Business intelligence
9min read
17 Sept 2026
A platform rebuild rarely fails on technology. It stalls on the parallel run, the undead source system and the decommission that nobody owns.
Data & AI strategy
7min read
15 Sept 2026
An agent's blast radius is set by its tool contracts, not its prompt. How to scope authority, gate writes, and keep the boundary testable rather than hopeful.
Artificial intelligence
7min read
10 Sept 2026
You own the platform and the budget, and every part of it is failing at once. A sequencing argument for what to fix first, what to buy, and what to defer.
Data & AI strategy
8min read
8 Sept 2026
Wrong measures do not crash. They return a plausible number every morning until someone finally checks. How they break, and how to catch them early.
Business intelligence
7min read
3 Sept 2026
Scheduled runs hide the flaws a backfill exposes. A practitioner's guide to idempotent writes, partition scoping, and proving that a re-run changed nothing.
Data engineering
7min read
28 Aug 2026
Autonomy is a cost, not a feature. Most tasks that reach for an agent are better served by a fixed sequence of steps you can actually test and reason about.
Artificial intelligence
4min read
24 Aug 2026
Token spend is the number everyone quotes and the one that matters least. The real cost of an agent in production is the evaluation harness around it.
Artificial intelligence
1min read
20 Aug 2026
Warehouse spend rarely spikes. It creeps, quarter after quarter, through small decisions nobody revisits. The fix is a review habit, not a bigger budget.
Data engineering
2min read