Most AI projects fail after the demo
The demo works because you asked the questions you had in mind. Production is a stranger asking something you never anticipated, and the model answering anyway. I work with teams on the part that comes after the demo: what the system is given, how you know the answer is right, and what happens when it is not.
The method
Reliability does not come from a better prompt. It comes from what the model can see when it is asked, and from what checks the answer before anyone acts on it. Four questions, in order, before anything is rebuilt.
Audit
A fixed-scope assessment of an AI system you already have, or one you are about to build.
I work through the system against the four questions above, plus the six ways these systems fail. You get a written account of where it will break, what it would take to fix, and which parts are not worth building at all. Most engagements start here, because it is the cheapest way to find out whether the rest is necessary.
Build
Embedded with your team, on the parts that decide whether the thing works.
Retrieval and context design, evaluation harnesses, the handling of uncertain cases, and the decisions about what a human sees before a customer does. I work alongside your engineers rather than delivering a system over the wall, so the capability stays with you when the engagement ends.
Fractional
Ongoing product leadership for a team without a senior AI product voice.
A recurring commitment where I hold the roadmap, the evaluation standard and the quality bar with you over months rather than weeks. Suited to companies where AI is central to the product and the decisions keep coming.
Advisory calls
A single conversation, when a conversation is what you need.
For founders and teams facing one specific decision: whether to build or buy, whether an approach will hold, why something in production is behaving the way it is. No proposal, no engagement, just the question worked through properly.
How engagements are priced
Project-based and scoped per client. What a system needs depends entirely on what it does and who depends on it, so a rate card would tell you almost nothing useful. The first call establishes what the problem actually is, and I follow it with a written scope and a fixed price before any work begins.