Any layout, any condition
Scans, phone photos, exports, born-digital PDFs. Rotated, stamped, multi-column, two languages on one page. The messy ones are the reason this exists.
Invoices, forms, statements, contracts, IDs — scanned, photographed or born digital. Arakma reads the fields you define, scores every one, and routes anything it is not sure about to a person before it reaches your system.
On your own documents · billed per page · stored encrypted in the EU
Arabic, French and English — including two on one page
What you get
Any model can read a clean page. The work is the one that is skewed, stamped, half handwritten, or missing the field you needed.
Scans, phone photos, exports, born-digital PDFs. Rotated, stamped, multi-column, two languages on one page. The messy ones are the reason this exists.
You define the schema once. Every response comes back with exactly those keys, typed and in that shape — not a bag of whatever happened to be on the page.
Calibrated per field, not per document, so one hard field does not drag a whole page down. You set the line — 85% by default — that decides what can go through untouched.
Anything under the line is checked by a reviewer with the document open beside the field. The correction is logged against the job, and it is how the system gets better.
The mechanism
Nothing uncertain is quietly guessed at, and nothing uncertain reaches you unchecked. You move the line; 85% is where it starts.
Delivered straight through. Nobody on your team sees it, because there was nothing for them to decide.
Held back and shown to a reviewer with the document open beside the field, then corrected before anything is sent.
One webhook, one payload shape, and every correction logged against the job with the actor and the timestamp.
The integration
Name the fields in a sentence or point at a saved schema. Post the file, get a job id back immediately, and collect the finished result on your webhook — after review, if review was needed.
curl https://api.arakma.com/v1/extractions \-H "Authorization: Bearer $ARAKMA_KEY" \-F "file=@invoice_8842.pdf" \-F "prompt=invoice number, issue date, vendor, total" \-F "webhook=https://yourapp.com/hooks/arakma"# Returns a job id immediately. Nothing blocks on the extraction.
{"jobId": "job_01HQ8F2","status": "done","fields": [{ "key": "total_amount","label": "Total amount","value": "12460.00","confidence": 1.0,"page": 1 }]}
Document types
The schema is the product, not the document type. Describe the fields you need in plain language, or save a schema and reuse it — the same extraction and the same review step apply either way.
These are the shapes it is built around. If yours is not here, that does not make it a special case — bring one and we will show you the output before you pay anything.
By industry
The same schema and review step, set up around the documents and the checks each trade already runs.
Invoice F-2026-0418, Purchase order PO-7710, Delivery note BL-2291
Learn moreNational ID card, Proof of address, Bank statement · 3 months
Learn moreConstat amiable, Photos · 6 images, Repair estimate DV-5510
Learn moreBill of lading MAEU2261, Commercial invoice CI-3307, Packing list
Learn moreMaster services agreement, Annex B · pricing, Amendment no. 2
Learn moreHandling
Chosen per tenant when the account is created. Documents are stored encrypted in that region and are not copied out of it.
Your documents are not used to train any model, ours or a vendor's. The processing agreement with our model provider prohibits it.
Thirty days by default, or whatever you set. The file and its extracted values are deleted together, and you can delete a job sooner by API.
Tell us which documents you handle and we open an account for your team. Upload a PDF or a photo, say what you need out of it, and see exactly what comes back — including which fields we would not trust without a person checking them.
We reply within one business day.