AI-native development
Routing, constrained generation, and the gap between a model that works and a feature that ships.
2 of 2
prose generations that invented a statistic
Capabilities
The work page is what I have built. This is what I can build for you — and every claim on these four pages carries a measured number or a named failure, because the alternative is a list of adjectives.
There is a through-line, and it is worth saying out loud: all four are about systems that have to be correct when nobody is watching them. A model that fabricates a statistic, an agent that degrades into guessing, a defect that reaches a million recipients before a dashboard refreshes, one tenant quietly damaging every other tenant, a funnel stage with nobody in it — none of these announce themselves. They all look like success until someone asks the specific question that exposes them.
Knowing which question to ask is most of what I am selling.
Routing, constrained generation, and the gap between a model that works and a feature that ships.
2 of 2
prose generations that invented a statistic
Twenty-seven specialised agents, an approval queue, and four failures that taught me what a fleet needs.
275 vs 11
enrichment errors against successes, unnoticed for months
A billion messages a month for eight years, and what that volume forces you to build differently.
0.1% = 1,000,000
why volume replaces inspection with measurement
Building the platform itself — where one tenant's behaviour decides whether everyone else's mail arrives.
1 bad tenant
is enough to degrade delivery for every other tenant
One contact layer behind three products, and the funnel failures that looked healthy from the dashboard.
28 → 0
active sequences, and the number of people they enrolled
If you would rather see the finished thing than read about the method, the three systems are live and openable.
Tell me which of the four your problem sits in — or that it sits in none of them, which is also an answer worth having.
Get in touch