An inquiry in, a defensible answer out — with receipts.
These are real screens from the product, not illustrations. Each is captioned with exactly what the machine did.
1 · Parse — the customer's words become typed constraints
GRADE reads the inquiry and extracts a structured list of hard constraints: a particle-size limit as a number, a chemistry as a string, PFAS-free as a boolean. Every extracted value carries a verbatim quote from the inquiry. If the quote is not present in the text, the value is rejected — the model cannot invent a requirement.
2 · Screen — deterministic SQL, never a guess
The constraints become a database query. A maximum particle size is a WHERE clause; it is never compared by embedding similarity. A product that fails a stated constraint cannot survive. A product missing data on a constrained field is flagged unknown for a rep to check — never silently passed.
3 · Rank — the model only sees survivors
Only products that passed the deterministic screen are handed to the model for ranking and rationale. It orders them and explains the fit; it cannot resurrect a grade that failed a hard constraint. Every recommendation still traces back to the screen.
4 · Draft — a reply your rep approves
GRADE writes a reply the rep reviews, edits, and sends. Nothing sends automatically. When the rep records what the customer actually bought, that outcome feeds a measured-accuracy number — the honest version of "it learns," scoped to that customer's own data.