Writing
What we learned the expensive way.
Notes from operating our own products. Positions we hold because something went wrong once, not because they sound good in a deck.
- 017 min read
Microsoft Clarity on Shopify: What It Shows You, and What It Structurally Cannot
Microsoft Clarity is free, good, and missing one thing on a Shopify store: it does not know what anything is worth. Here is exactly where that boundary sits.
Microsoft ClarityShopifyAnalyticsSession replay - 026 min read
Every agent we ship has an approval gate
Autonomy is not the goal. An AI agent that can act on production data needs a bounded blast radius, and approval gates are what make clients say yes at all.
AI agentsProduction AIApproval gatesAgent architecture - 037 min read
If you did not run a holdout, you do not know it worked
Before-and-after numbers are not evidence. Why we build holdout measurement into conversion work, and what it costs to report a number you can actually defend.
AttributionHoldout testingIncrementalityConversion optimisation - 048 min read
When a smaller language model is the right call
Frontier models are the right default until cost, latency or data residency says otherwise. How to find out which case you are in, before committing.
Small language modelsSLMFine-tuningInference costModel selection
Working on something in this territory? We would rather talk about your actual problem than send you a newsletter.
Start a project