Notes from the build floor.
Engineering deep-dives, opinionated takes on what works in AI delivery, and the occasional "we tried this and it failed" essay.
How we rebuilt the ERP we wanted but couldn't buy.
Abstract started as a six-week prototype for one 3PL. Eighteen months in, we wrote down what we'd learned about building enterprise software you'd actually want to use.
Cost-routing prompts across providers (without the LangChain).
A tiny router that decides which model handles each prompt — by latency, cost, and an eval gate. 90 lines of Go. Six months in production.
Build, buy, or stop. The AI question we ask first.
Not every AI project ends up shipping. Here are four questions we work through with every team before writing a line of code.
Explainability isn't a feature. It's a requirement.
MindSight flags seven conditions. For every flag, the clinician sees the evidence, the confidence band, and the version of the model that decided it. Here's why.
Webhooks that don't lie. A pattern for at-least-once delivery.
Reconciliation, idempotency keys, conflict resolution. Everything we wish Shopify's docs would just say plainly.
Why forecast accuracy is the wrong target.
Ops teams chase 99% accuracy and miss the point. The right metric is "decisions improved" — and it's harder to measure but cheaper to act on.
Evals are the new tests. Stop treating them like dashboards.
Three patterns we use for production AI evals — including the one nobody talks about: the eval-gated deploy.
Why every UnknwnAI engagement starts with six weeks.
Not three months, not a year. Six weeks. Long enough to find the signal, short enough to fit inside a quarter, fixed-price.
Designing data rooms clinicians don't hate.
The audit trail is also a UI. What we changed in MindSight so that compliance felt like a feature, not a hurdle.