How it gets things right
Three mechanisms a buyer is entitled to ask about, described as they actually work. None of them is a number we can put on a page — they are disciplines, and the discipline is the reason to trust the output.
Quality
It is graded against quotes you already sent
The test set is not synthetic. It is real requests from your own inbox, scored against the proposal your team actually sent back — the only ground truth that counts, because it is what a customer accepted.
- Every case is re-run rather than remembered. Verdicts are recomputed against the live pipeline each time, never read back from a stored result.
- Runs happen with the learned layers deliberately emptied, so a score cannot be flattered by an answer the system was handed earlier.
- A change to matching is measured against the whole corpus before it ships. One that does not beat what is already there does not go out.
- Cost per request is measured in the same run as accuracy, so a more accurate pipeline that costs ten times more is a visible trade rather than a surprise on an invoice.
Completeness
It proposes the parts that go with the part
A luminaire needs a driver. A run of profile needs connectors and end caps. Customers rarely list them, and a quote that omits them comes back — or worse, ships incomplete. Ettore derives them from the rules of your catalog and computes the quantity.
- Each proposal lands as one of four things: a required line that blocks delivery until it is resolved, an optional line the rep decides on, a flag that warns without adding anything, or a plain statement that nothing needs adding and why.
- Quantities are computed rather than guessed — a formula over the line’s own facts, not a rule of thumb.
- When a fact it needs is missing, it abstains and says so. Unknown is never treated as zero.
- The rules are written against your catalog. Where you publish a price table carrying dependency rules your ERP does not hold, that is where they come from.
Memory
It becomes where the answers live
The knowledge that currently sits in one person’s head — which competitor code maps to which of your products, which spec is right when two sources disagree — accumulates in one place instead. That is the part that compounds.
- A confirmed match is remembered. Next time that reference arrives, from any customer, it resolves immediately instead of going through matching again.
- Equivalences mined from your own history arrive as suggestions, never as facts. Nothing learned changes an answer until a person activates it.
- Where two sources disagree about a spec, both are kept side by side and the conflict is queued for someone to settle — rather than one silently overwriting the other.
- Every correction a rep makes is an entry. The work of fixing a match is never thrown away.
No accuracy figure appears on this page. Any single number would need so much context to be honest that it would stop being informative — ask us for the eval report against your own requests instead.