asjson

schema in, json out,
a probability on every field.

Send a JSON Schema and a document. Get back a record that matches the schema, with a calibrated confidence you can threshold on every value.

you send

schema
{
  "properties": {
    "intent":   { "enum": ["refund","cancel"] },
    "order_id": { "type": "string" },
    "urgent":   { "enum": ["yes","no"] },
    "amount":   { "type": "number",
                  "nullable": true }
  }
}

document
"I still haven't had the $48 back on
 order 41822. Third time I've written."

you get

intent"refund" 0.96
order_id"41822" 0.99
urgent"yes" 0.88
amount48 0.61

Every value carries its own probability, and the evidence span it came from. Route the confident ones, queue the rest.

why not just an llm

No token streamThe record is decided, not written. There is nothing to parse and nothing to retry.
No parse errorsEvery enum is one of the listed options, every array is in bounds, every number is a number. The response matches the schema you sent.
Confidence you can thresholdEach field comes back as a distribution, not a string. Set a bar, act above it, review below it.
Dependent fields in one passA chosen tool gates its own arguments. The document is read once.

what people point it at

Tool routing

Turn a sentence into a call. The tool and its arguments come back together, each with its own confidence.
Triage and extraction

Pull the fields you need out of a ticket, a filing or a thread, and let the low-confidence ones fall to a human.
Judging and QA gates

Score an output against a rubric you define as a schema, and gate on the number rather than on a sentence.

schema in, json out,a probability on every field.

you send

you get

why not just an llm

what people point it at

Tool routing

Triage and extraction

Judging and QA gates

early access

schema in, json out,
a probability on every field.