Applied AI Lab
A supervised-autonomy pipeline on AWS — the model drafts, a deterministic gate recomputes every figure it quoted, a human signs off, and only then does anything publish.
The problem. Publishing anything with numbers in it — a portfolio report, a risk commentary — is the class of work where a plausible sentence is worse than no sentence. A language model is very good at the prose and structurally incapable of guaranteeing the arithmetic. Most products solve this by asking the model to be careful.
What runs. A Step Functions workflow with three rungs. The model drafts commentary and must return, alongside the prose, a declaration of every figure it used. A gate with no model in it recomputes each of those figures from the source data and rejects any number that is not there; a failure loops back to drafting with concrete feedback, three strikes and the run dies without shipping. What survives goes to a human on an out-of-band approval step and waits — indefinitely, by design. A second module applies the same contract to volatility and risk analytics. Everything runs on synthetic data in a personal account.
Designed next. Natural language in, verified evidence out: a compiler from a human sentence to a whitelisted, parameterised query that is shown before it executes — so the contract the system will honour is visible while it can still be corrected. Same discipline pointed at questions instead of reports.
The transferable part. Invention is allowed in the prose and never in the figures. That single line decides the architecture: the gate cannot contain a model, and autonomy is earned one rung at a time rather than granted.