Pricing
Fixed scope. Objective acceptance. No surprise invoice.
Each tier has an acceptance test agreed before work starts — a checkable number, not a promise. The guarantee is never weakened to hit it.
Each tier is a one-time, fixed-scope engagement — not a subscription. You own what’s delivered; Production adds 30 days of support after handover.
Pilot
Prove it works on your data.
$1,200
One document type (up to 12 fields).
- Schema design for one document type (up to 12 fields)
- Extraction pipeline: ingest → extract → validate → confidence → structured output
- Accuracy report on a held-out slice of your documents: field-level exact match, per-field precision/recall, false-invention rate
- JSON + CSV output, CLI, documentation
Acceptance test
The accuracy report runs on your held-out documents and reproduces on your machine. Objective, checkable, agreed before work starts.
Certified pipeline
The guarantee and the review loop.
$3,000
Everything in Pilot, plus certification.
- Everything in Pilot
- Calibration certificate: certified error bound α on auto-accepted fields, at your chosen risk level
- Two-pass extraction (Hunter + Mapper) with cross-check
- Deterministic provenance: page and position per field
- Human-review queue with corrections captured and fed back as few-shot examples
- Validation rules, Docker + tests
Acceptance test
The certificate is issued at the agreed α with observed selective risk at or below it on held-out data. The guarantee is never weakened to hit a number.
Production service
Run it in production.
$6,500
Everything in Certified, up to 3 document types.
- Everything in Certified, up to 3 document types
- FastAPI async service
- CI/CD accuracy-regression gate
- Drift auto-de-certification
- PII / redaction / retention
- Handover runbook, 30 days support
Acceptance test
The service meets the agreed certificate on held-out data, and the CI regression gate holds the guarantee on every change.
What acceptance means
You agree the test before you pay.
The number is set up front
For a certified pipeline, the certificate is issued at the agreed risk level α, with observed selective risk at or below it on held-out data. If your documents will not support that, the pipeline says so and routes everything to review rather than pretending.
It reproduces on your machine
The accuracy report runs on your held-out documents and reproduces on your hardware. Objective and checkable — not a slide.
Labelled data is a hard requirement
Without labelled samples the pipeline refuses to auto-accept. This is not a limitation to work around — it is the reason the number means anything.
Not sure which tier? Start with a pilot.
One document type, an accuracy report on your data, five business days. It tells you whether certification is worth it before you commit to it.