Blog
Practical guidance on AI adoption, cloud economics, data privacy and software delivery — written by the engineers doing the work, and specific enough to act on.
Most AI proposals start with a model and end at a data problem. Here is the assessment that runs before either, and what a genuinely ready institution looks like.
Read the articleMost AI proposals start with a model and end at a data problem. Here is the assessment that runs before either, and what a genuinely ready institution looks like.
Read
Accuracy is the easy part. Being able to explain where each extracted value came from is what makes it defensible.
Read
Scoring at intake catches different fraud from scoring at settlement. Most teams invest in one and miss the other entirely.
Read
Deterministic matching on phone, email and loyalty number beats a learned model for most retailers, and it can be explained.
Read
Egress, engineering time and the deferred-remediation line item. A realistic migration model needs all three, or the first invoice will be a surprise.
Read
The order you move workloads in determines whether a failure is an incident or an outage. A defensible wave plan starts with the lowest-risk target.
Read
Every vendor says you are not locked in. Here is how to test that claim yourself, in an afternoon, using only your own systems.
Read
Replacing the core is rarely the answer. Screen-scraping is never the answer. What works is a narrow, contract-tested adapter layer.
Read
Accessibility retrofitted before launch costs more and performs worse. What changes when it is a build criterion instead of a compliance gate.
Read
Bad signal, gloves, no patience. Conflict resolution is where offline-first implementations actually fail, and it is a specification problem.
Read
Banks and insurers do not get to restart. Versioning strategy matters more when the system has a decade-long service life.
Read
Alert fatigue is a design failure. Most teams instrument dozens of signals and still discover outages from customers.
Read
The usual answer focuses on licence cost. The more useful question is how often your process will differ from everyone else's.
Read
You do not need 90% coverage to make a system safer. Characterisation tests around the parts that matter buy more per hour.
Read
The technical cause is the least interesting part of the update. A structure that keeps people informed without alarming them.
Read
Specs fail because they describe screens rather than behaviours. A practical structure for acceptance criteria that outlive the first release.
Read
A component library without documented usage rules becomes an inconsistent library within a year. What to write down so the system holds.
Read
Refactoring appeals to engineers and repays nothing on its own. Framing the same work in delivery risk terms changes the conversation.
Read
If your ML system cannot produce a defensible explanation for a decision, the governance framework is decorative. Here is what to keep.
Read
Consent is a data-model property before it is a legal one. How to build retention and purpose limits into the schema rather than bolt them on.
Read
A model that is accurate and ignored saves nothing. The failure mode is organisational, not technical, and it is predictable in advance.
Read
Store teams re-keying customers is an adoption failure, not a data integration failure. The technical work is the easy half.
Read
A screen showing on-time performance tells a dispatcher nothing. Visibility is exception-first and arrives before the customer complains.
Read
Get in touch
We would rather answer a specific question than pitch. If you are working through something similar, send it to us and you will get a considered reply from an engineer.
Understand what we think first, then decide whether a conversation is worth your time. Either order works.
23 articles across 5 categories, most recently published 2 October 2026.
We would like to use Google Analytics to understand how this website is used. No analytics are loaded unless you accept. Your choice is stored for six months.
See our Privacy Policy for details.