Applied AI, held to a number
Classifiers and assistants that are scored rather than demoed. The feedback classifier triages incoming product feedback at an F1 of 0.98 against a labelled set. The chatbot authors its conversation paths as an explicit decision tree instead of leaving them to the model, with Application Insights wired in so the paths people actually take are measurable rather than assumed. The interesting part of this work is never the model — it is the labelled set, the failure cases, and the honest number.
Process work
Finding where a process actually costs time and money, then modelling the change before anyone commits to it. The Business Case Tool exists to do exactly this for its users: prospects configure who owns each step of their process, and the tool models the time and cost that moves. Brood applied the same instinct internally — sourcing and voting on hackathon ideas ahead of the event, behind single sign-on so participating cost nobody an account. The useful output is usually a smaller system, not a bigger one.
The correctness work underneath
Both of the above fall over without it. For 2K Build Lab that meant rebuilding a shipped game’s rating engine from its binary tuning file and holding the port to a byte-identical parity standard, because an optimizer that returns a plausible answer is worth nothing. The Vue Devtools Auditor comes from the same instinct: report an accessibility violation at the component that owns it, not at a line of rendered output.
How engagements work
There is no rate card here, because I have never had one that survived contact with a real scope. Tell me the problem and what a correct answer would look like, and I will tell you whether I am the right person and what it would take. If I am not, I will say so.