Product

One assay stack for every stubborn file.

Parsing, splitting, extraction, and a studio so domain experts can iterate in the same week they find a new packet type.

Vision
Recover layout, reading order, tables, handwriting. Thirty-plus file families land as blocks with coordinates.
Language
Map recovered structure onto your schema. Types, allowed empties, and nested objects stay in the mold.
Correction
Last-mile checks against the source page. A reviewer sees the span that justified each value.
Serving
Batch, realtime, and private-cloud inference. Cloud, VPC, and air-gap share one API surface.

Parse

Reading order that matches the page.

Columns, footnotes, rotated scans, nested headers. Parse emits structured blocks with coordinates so a later step can cite the cell. Markdown is a convenience view. JSON is the contract.

  • Thirty-plus file families, including PDF, TIFF, PNG, JPEG, DOCX, PPTX, XLSX, HTML, and email wrappers.
  • Table reconstruction with merged cells and multi-row headers.
  • Figure captions bound to the drawing they describe.
  • Spreadsheet sheets as first-class objects, billed per sheet.
Parse API

Extract

A schema is a mold. Pour only what fits.

Define fields, types, and allowed empties. Cupelin returns JSON plus the spans that justify each value. Low-confidence fields can pause for review. Confidence chips in Studio use bead gold above 0.85 and kiln below that floor.

Idempotency keys on create. Webhooks when a long job cools. Same schema in Studio and in POST /v1/extract.

Extract API

Split

One upload. The documents inside it.

Describe sections in plain language. Get page ranges back. Run extract per child document. A 90-page claims binder becomes a FNOL, a medical packet, photos, and a coverage letter, each with its own schema.

Split is a first-class job. Child file IDs stay stable across retries so a night batch can resume on the packet that failed.

Binder recipe
FNOL · pages 1 to 4 · extract claim_intake
Medical · pages 5 to 22 · extract cms_1500
Photos · pages 23 to 31 · parse only
Coverage letter · pages 32 to 36 · extract policy_terms

Studio

Fire small batches before you fire the warehouse.

Upload examples, lock a schema, run evals, compare versions. Studio is for the people who know the document. Ten seats on Foundry. Two on Assay. Private fire includes a named TAM who sits in the same canvas.

Use cases

Where messy paper still runs the business.

01

Contracts and diligence

Clauses, dates, parties, exhibits. A fund ops team can cite the page a change-of-control sat on.

Open the recipe
02

Invoices and AP

Line items that survive multi-page tables. Totals stay tied to the cell they were read from.

Open the recipe
03

Insurance and claims

Packets, photos, handwritten notes. Split the binder, then extract each child with its own mold.

Open the recipe
04

Research corpora

Papers and filings into retrieval that can cite a table. Figure captions travel with the drawing. A later agent can point at the span.

Open the recipe
05

Agent tool use

Give an agent a parse it can point at. Cupelin returns blocks, boxes, and file IDs. The agent never executes a real-world action through us.

Open the recipe

Customers

Built for teams that live in PDFs.

High-volume, layout-heavy, often regulated documents. If that is your week, Cupelin is in your category.

Document API

A parse-and-extract layer inside a larger legal product. Citations stay on the clause so a lawyer can audit the model’s keep.

Fund ops

A residency-bound team on Private fire. Air-gap included. Named TAM. Volume billed on a 2,000,000 page floor.

Insurance intake

A desk drowning in packets. Split first, extract second, review the kiln-chip fields before a claim moves.

Talk to us about volume and residency