Insights
Practical notes from real AI deployments. What works, what stalls, and what to do first. No hype.
The same failure mode I had to design around building a CV parser shows up, at much higher stakes, in how electronic health records move between systems. What silent data loss actually looks like, and the nothing-dropped rule that fixes it.
Building software for places where the internet is a maybe changes how you design everything, not just the sync button. Patterns from real field work, no client names, just what actually holds up.
A year of building with AI doing most of the typing. What it speeds up, what it quietly costs you, and the METR study that explains why it can slow down your best people on your hardest codebases.
Working AI in 30 days and governed production in 90 is a real promise, not a marketing number. Here's exactly what lives inside each window, and what I cut when a client tries to rush it.
NIST AI RMF and ISO 42001 have real, published timelines, and none of them fit in 30 days. Here's what the certification bodies and NIST itself actually say, so you can call out the vendors who are lying to you.
New 2026 CPT codes removed the thresholds that made RPM impractical for small practices. What changed, what it pays, and how to run a program without burning out your staff.
The verified failure numbers behind enterprise AI pilots are brutal, and the reasons are mostly boring. What actually gets working software into governed production.
