Document AI

PDFs in.
Fields out.
Nothing leaves Nairobi.

Stop re-keying passport numbers, policy clauses, and claim amounts from scanned stacks, extract on private models with audit trails your compliance team can defend.

$ corvux doc extract --batch claims-q3 \
 --schema motor-claim-v2

INTAKE your PDF pile, counted on the job
CLASSIFY police_ob · garage_inv · id_copy · photos
EXTRACT reg_no · incident_date · repair_total · policy_ref
VALIDATE policy lookup API · human on low confidence
REVIEW low-confidence → human queue
✓ write-back: claims_core · host kenya-gpu-example

Kenyan back offices drown in PDFs: KYC packs from branch scanners, lease agreements with handwritten riders, motor claims with police abstracts and garage invoices stapled in random order. Someone re-types every field into core banking or claims systems, slowly, expensively, and with typos that become disputes.

Document intelligence at Corvux is extraction infrastructure, real extraction, free of a demo OCR sticker. Classify pages, pull structured fields, validate against your APIs, and route low-confidence rows to humans, all on models running in Kenya.

Schema
Your fields
Review
Human gate
DMS
Write-back
KE
Inference

Where it lands first

Queues with expensive keystrokes.

Bank KYC

ID copies, proof of address, CR12 extracts, flag missing pages before a compliance officer opens the folder. Pair with banking workflows.

Insurance claims

Motor, medical, and property packs, summarise narrative, extract amounts, match policy numbers. See insurance.

Contracts & leases

Commercial terms, renewal dates, penalty clauses, structured output for legal ops and procurement.

Board & EXCO papers

Feed downstream to meeting action engine for minutes and action tracking.

Pipeline shape

Ingest → understand → validate → approve.

01
Schema workshop
Define fields, validation rules, and “never guess” thresholds with your ops owner.
02
Sample benchmark
200–500 real documents (redacted). Measure precision/recall per field before production.
03
Integrate
Watch folders, email inboxes, or API upload → extraction → your DMS/claims/core system.
04
Operate
Drift monitoring when new form layouts appear; retrain triggers via fine-tuning.
The useful number is yours: how long a clerk spends per file today, and how many rows a person still has to touch after the model drafts. We measure that after we see the scans. No invented 18-to-4.

Stack with AI automation for full queue orchestration, or managed sovereign AI if you need GPU capacity without buying hardware.

FAQ

From ops and risk.

Does this replace our OCR vendor?

Often we sit downstream: OCR gives text, we classify pages, extract entities, and validate against your schema. Sometimes one pipeline replaces both.

How accurate on Kenyan ID layouts?

We benchmark on your sample set, old ID cards, new e-passports, CR12 forms. Pilots ship with field-level confidence scores and human review queues.

Can it read handwritten doctor notes on claims?

Mixed. Typed hospital discharge summaries yes; illegible scribbles get flagged for manual review rather than guessed.

Where does extraction run?

Kenya GPUs via managed sovereign AI or your on-prem cluster. No document bytes sent to OpenAI.

Pilot extraction on one document type

Bring 200 sample PDFs and your target schema. We return benchmark numbers and a residency architecture.

Book a document AI workshop