Industry · KYC, identity verification, fraud, document AI
Identity and KYC training data with a consent record your counsel can audit
Face and liveness models trained on Western datasets fail on South Asian skin tones, lighting and devices, and document OCR trained on Latin or Arabic fails on Nastaliq. We collect consented biometric capture, presentation attacks and Urdu-native document ground truth — with per-record consent designed for biometric-privacy law.
Who this is for
For teams verifying South Asian users, wherever those users are
Pakistan has 240 million CNIC-eligible citizens and a fintech sector onboarding millions a year; the diaspora onboards through UK, Gulf and North American banks.
- KYC and IDV vendors — face match, liveness and document verification models that must work on Pakistani and South Asian users
- Banks, wallets and fintechs — onboarding and step-up authentication for Pakistan's mobile-money users
- Fraud and anti-spoofing teams — presentation-attack data: print, replay, 2D and 3D masks under realistic conditions
- Document AI and RPA — Nastaliq OCR for CNIC, utility bills, bank forms, prescriptions and ledgers
- Global platforms — skin-tone and demographic diversity for fairness audits across face pipelines
What breaks
Where identity pipelines break on South Asian users
The failures are demographic, environmental and orthographic.
Skin tone and demographic coverage
Face pipelines trained on North American and European faces show higher false-reject rates on South Asian users. Consented, quota-balanced capture across Punjabi, Sindhi, Pashtun, Baloch and Muhajir subjects fixes the distribution — and documents it for fairness audits.
Presentation attacks in the wild
Print, screen-replay and mask attacks look different under Karachi fluorescent light on a budget Android camera. Attack data has to be captured under the same conditions as genuine samples.
Nastaliq documents
CNIC, bills and forms mix Nastaliq Urdu, English and numerals. OCR trained on Naskh Arabic or Latin cannot segment it. See why Urdu OCR is hard.
Consent that survives legal review
Biometric data without evidenced informed consent is unusable under Illinois BIPA, GDPR Article 9 and similar regimes. Every subject in our sets signs a plain-language biometric consent, versioned and delivered per record.
What we supply
Consented biometrics and document ground truth
Field capture in six cities, consent designed for biometric-privacy law, delivery to the region your compliance requires.
Face and liveness capture
Multi-pose, multi-lighting, multi-device capture of consented subjects, balanced across ethnicity, age, gender and skin tone, with matched print, replay, 2D and 3D mask presentation attacks. See the Face & Liveness Set.
Document OCR ground truth
Printed Nastaliq and handwritten forms, receipts, prescriptions and ledgers with reading order, field-level labels and layout boxes; synthetic-free, real documents with PII handled per spec. See Nastaliq & Handwriting OCR.
Voice biometrics
Enrolment and verification utterances across sessions and devices in Urdu, Punjabi and Pashto, with matched replay attacks, for speaker verification.
Annotation on your data
Face landmarking, document field extraction, fraud labelling and fairness-audit labelling by trained annotators, remotely or inside your perimeter.
Ready-made
Catalog datasets for identity and KYC
Each ships with a datasheet, per-record consent artefacts and a free unwatermarked sample pack.
Pakistani Face & Liveness Set
Consented biometrics, presentation attacks
ImageNastaliq & Handwriting OCR
Forms, receipts, prescriptions, ledgers
SpeechPakistani Telephony Corpus
Speaker verification source audio
ImageKiryana & Informal Retail Imagery
Real-world capture conditions
TrustConsent & fair pay
How biometric consent is obtained
TrustSecurity
Encryption, access control, secure facilities
How a programme runs
From compliance review to delivery
Identity buyers start with legal. So do we.
Consent and DPA review
Consent form, licence text and DPA sent within one business day; your counsel reviews before any capture.
Spec the distribution
Ethnicity, age, gender, skin tone, device, lighting and attack types, with quota tolerances and an acceptance test.
Pilot capture
500–1,000 subjects or 10k–50k pages, QA-passed, delivered to your region for evaluation.
Scale with audit trail
Per-record consent artefacts, subject pseudonyms and capture metadata on every delivery, so any row traces to its signature.
Answers
Questions identity and KYC teams ask
Each subject signs a plain-language biometric consent that names machine-learning training, storage and sharing, and the consent artefact is delivered with the record. That is the evidence BIPA and GDPR Article 9 require. Your counsel decides applicability; we provide the documentation.
Yes. Print, screen replay, 2D and 3D mask attacks captured under specified devices and lighting, matched to genuine samples from the same subjects, and iterated against your model's failure cases.
We produce OCR ground truth on document types you specify using consented, real documents with PII handled per your spec, or on your own document images inside your perimeter. We do not source or resell citizens' ID documents.
Delivery is to the cloud region you choose; on-prem and in-VPC annotation are available; retention and deletion follow the contract.
Yes — most of our licensing goes to AI teams outside Pakistan. English-language MSA and DPA, USD invoicing by default, W-8BEN-E for US payers, delivery to your AWS, GCP or Azure region. See Where we deliver for compliance mapping by region.
Failing on South Asian faces or Nastaliq documents?
Ask for the consent pack and a sample. Both come back within one business day.
Average first response: under 6 business hours. NDAs signed same day.