Document AI that never leaves your network.
Ready-made pipelines to read, classify and extract data from the documents your business runs on, deployed on Gerniq inside your own environment.
Fast models first, a vision-language model only where it's needed.
Most pages are handled by efficient CPU models. Hard or low-confidence pages go to a vision-language model on GPU, which keeps both accuracy and cost under control.
Every step is a versioned model in the registry, so a model update never silently changes your results.
Starting with the workflows that hurt most.
KYC and onboarding
Identity documents, address proofs and application forms. Extract names, numbers, dates and addresses, and flag mismatches between documents for review.
Loan and credit files
Bank statements, salary slips, tax returns and property papers. Split long files into document types and extract the figures underwriters check.
Insurance claims
Claim forms, hospital bills, discharge summaries and invoices. Extract policy numbers, line items, amounts and diagnoses.
Healthcare records
Lab reports, prescriptions and referral letters. Digitise inside the hospital network, where patient data must stay.
What makes it more than an OCR demo.
Extraction is the easy part. Running it on millions of pages, reliably and auditably, is what Gerniq is for.
- Confidence on every field, so you can route only uncertain documents to people
- Page-level parallelism for files with hundreds of pages
- Async jobs with retries and dead-letter handling for large batches
- Versioned pipelines with a clear record of which model produced each result
- Your own models alongside ours, fine-tuned or trained on your documents
- Metrics per step: throughput, latency and error rate for each model
How we measure a pilot.
We agree targets with you before the pilot starts, measure them on your documents and your hardware, and report weekly.
| Metric | How it is measured |
|---|---|
| Field accuracy | Against a labelled sample of your own documents |
| Straight-through rate | Share of documents that need no human correction |
| Cost per 1,000 pages | Infrastructure cost on your hardware or cloud account |
| Latency | p95 per page for synchronous calls; end-to-end time for batch jobs |
Bring 50 of your documents.
On a scoping call we will pick one workflow, agree the targets above and plan a 6 to 8 week pilot in your environment.