Turn an expert bottleneck into a self-serve product
A deterministic engine gets the numbers right, an AI varies the plan, and the expert reviews before it reaches the customer.
Precision Physiques is built on this — deployed, coach-exercised, pre-first-client.
A Mifflin-St Jeor macro engine computes each client's numbers deterministically. Gemini varies the five-day meal plan around those numbers. The coach reviews and approves every plan before the client sees it, delivered as a branded PDF.
Tell MAST which piece of work consumes your week. Half an hour is enough to see whether an agentic system is worth building.
Start a conversationFour parts, one guarantee: the targets are never left to the model
- Engine
Deterministic core
The maths that must be right — every time — runs in ordinary code, not a model call.
- Variation
AI-generated plan
The model varies the output within that guaranteed frame, with a deterministic fallback if it fails.
- Sign-off
Expert approves
The qualified person reviews before the customer ever sees it — always in the loop.
- Delivery
Async + branded PDF
Generation runs as a background job, not a spinner, and lands as a document the customer keeps.
Any expert bottleneck with a computable core fits this pattern
Risk-profile intake → deterministic allocation model → adviser-reviewed recommendation, instead of a manual spreadsheet per client.
Job inputs → a rules engine sizes and prices the spec, an expert signs off before it goes to the customer.
Anywhere a qualified person currently hand-builds a personalised output, this pattern turns it into software with the person still on the sign-off.
Does the AI just make things up?
No. A deterministic engine computes the numbers that matter (the ones that must be right every time); the AI varies the food and portions inside a frame the engine checks, and it has a deterministic fallback if it fails.
Where does the expert fit in?
In the loop, not replaced. The system drafts; the qualified person reviews and approves before the customer sees it — the pattern proven at Precision Physiques.
What if generation is slow?
It runs as an async job — the customer isn't stuck on a spinner waiting on a model call. This replaced a ~58-second synchronous wait in the live build.
Tell MAST which piece of work consumes your week. Half an hour is enough to see whether an agentic system is worth building.
Start a conversation