AI products · Local prototype
IntentCart
Plan a setup. Check the evidence. Keep approval yours.
A local planning and spending prototype that turns reviewed requirements into a compatible fictional-catalog basket, exposes missing evidence and requires approval of the exact quote.
Recorded prototype
Recorded local prototype · fictional catalog
The question
A plausible recommendation can still have missing compatibility facts, a stale price or no permission to buy. Those are different questions and need separate checks.
How it works
- Represent requirements, catalog facts and compatibility constraints explicitly; defer when the available evidence cannot support a basket.
- Separate feasibility, finite-catalog optimality, quote approval and execution status instead of showing a single reassuring badge.
- Use exact quote approval, offline failure replay and fake-provider reconciliation; make the plan inspectable in a 3D workspace.
Evidence & results
- Saved evaluation covers 384 fictional goals: 73 feasible minimum-cost plans and 311 correct deferrals against a finite reference oracle.
- The local targeted inspector used 143 fixture lookups versus 1,920 for a full census in that evaluation.
- Existing media shows the actual local interface. Historical sandbox receipts are separate from the current bundle and from simulated recovery tests.
Scope & next steps
- Fictional catalog and authored scenarios; these results do not establish real shopping outcomes, model superiority or customer savings.
- The live-model comparison is dry-run/contract-tested only. The CPU imitation learner did not beat the practical decision-splitting baseline.
- No public deployment or production payment service is established.
Go one level deeper.
Ask about the implementation, tradeoffs, or my contribution.
Explore with the portfolio guide