No thought leadership, no predictions for next year. Field notes from real projects, written by the engineers who were there: what we built, what broke, and what we'd do differently, including the parts that were embarrassing.
Architecture
Featured
Running a new settlement system alongside a twenty-year-old mainframe costs roughly double the infrastructure and two engineers watching a reconciliation dashboard. We were pushed to cut it to four weeks at the halfway point. Three of the nineteen discrepancies we found surfaced after week eight, in month-end processing — which, by definition, only happens monthly.
AI
How we build golden datasets that catch regressions before customers do, why 200 hand-labelled examples beat 20,000 synthetic ones, and the retrieval failure that only appeared when we sorted by document age.
Hiring
Three years of data from replacing whiteboard rounds with paid work samples — acceptance rates, first-year attrition, and the demographic shift we didn't set out to cause but are glad about.
Craft
We had a sensible technical sequence that would have shown clinicians nothing usable for five months. One conversation in week three changed the plan, and the project succeeded because of it.
Cloud
A FinOps write-up from three migrations. The savings almost never come from rightsizing instances — they come from finding the four workloads nobody could name an owner for.
AI
A model that was 6% better on paper and ignored in practice. What a week of ride-alongs found that two years of telemetry hadn't, and why we now budget for fieldwork.
Architecture
We've talked four clients out of rewrites in the last two years. Three were right to stay put. Here's the framework we use, and the one case where rewriting genuinely was correct.
Craft
Automated checks caught 31 issues. A two-hour session with four screen-reader users caught eleven more, and every one of them was the kind that makes a product unusable rather than merely annoying.
Teams
An operations team had twenty years of intuition about a batch system that simply didn't transfer to an event-sourced one. What we got wrong, and how we estimate handover now.
Architecture decisions we got right, and the ones we had to undo. No product announcements, no webinar invitations. Unsubscribe in one click.
Tell us the shape of the problem. We'll come back with a build plan, a team shape and a number — no obligation, no deck.