Identity systems / Ooredoo Maldives
AI-Driven eKYC
Building the intelligence behind digital identity verification.
- The problem
- Build identity verification in-house, with less reliance on vendor licensing.
- My contribution
- Architecture, model integration and liveness detection.
- The outcome
- Document processing reduced from about 30 seconds to 1–2 seconds; automated provisioning rose from 13.1% to 91.7%.
In this case study
Overview
Digital onboarding needs to connect a document to a person, then pass a usable verification result into a service journey. I built the core AI eKYC pipeline used in Ooredoo Maldives Digital SIM Activation, combining document detection, OCR, face matching and liveness checks.
My role
I architected and developed the core AI verification pipeline for Ooredoo Maldives Digital SIM Activation, fine-tuned and optimised open-source models, and implemented liveness and spoof checks. My work covered system design, model integration, testing and evaluation.
Problem
The task was to develop an identity-verification capability in-house and reduce reliance on vendor licensing. Reading a document is only one part of that task: the system also needs to match the person to the document and check for liveness before a verification result can inform onboarding.
Approach
Separate the verification stages
Organised document detection, field extraction, face matching and liveness as distinct responsibilities. This makes the system easier to reason about than a single, opaque verification result.
Adapt existing models
Used YOLOv8, PaddleOCR and InsightFace as the core model stack. Fine-tuning and optimisation focused on the identity-verification task rather than training every component from scratch.
Include liveness in the solution
Implemented liveness detection alongside the document and face checks. A successful face match alone does not describe the whole verification process.
Architecture
The responsibilities inside the verification journey.
AI-driven eKYC
A document.
A person.
A clear next step.
- Input
- ID / Passport + selfie
- Output
- Verification decision
- 01Document scan
- 02Face & presence
- 03Verification result
Explore the technical sequence
- ID / PassportDocument input
- Document detectionYOLOv8
- OCR / Field extractionPaddleOCR
- Face detection / MatchingInsightFace
- LivenessPresence check
- DecisionVerification result
- ProvisioningService hand-off
Challenges
Connecting different model outputs
Detection, extracted fields and face matching solve different parts of the problem. The architecture keeps their responsibilities explicit so the verification result can be interpreted in context.
Measuring the right boundary
Document-processing time and the full KYC journey measure different things. The system also has to complete face matching, liveness checks and the service hand-off.
Outcome
After rollout and optimisation, document processing fell from about 30 seconds to 1–2 seconds and automated provisioning increased from 13.1% to 91.7%. The deployed process later achieved 90% instant approvals. The in-house system reduced reliance on third-party eKYC licensing and manual verification. I also showcased the solution at MWC Doha 2025.
Reported results & measurement context
- ~30s → 1–2sDocument processing
- CV-reported reduction after rollout and optimisation. This measures document processing, not total KYC journey time; no separate combined AI verification latency is reported.
- 13.1% → 91.7%Automated provisioning
- CV-reported change after rollout and optimisation. This is a provisioning outcome, not a face-matching or OCR accuracy score.
- 90%Instant approvals
- Reported for the deployed process at a later stage. This is an approval outcome, not model accuracy, and is separate from the automated-provisioning measure.
Three different clocks
The 1–2 second figure covers document processing. It does not describe total KYC journey time.
- Document-processing latency
- Reduced from about 30 seconds to 1–2 seconds after rollout and optimisation, as reported in the updated CV.
- AI verification latency
- Document, face and liveness checks together. No separately scoped measurement is documented.
- Total KYC journey time
- Capture, uploads, verification, decisions, retries and service hand-off. No end-to-end measurement is documented.
Source: Yoosuf Sayyid’s CV, selected technical and delivery projects. Figures are self-reported; no independent benchmark or analytics report is linked.
Technology
- YOLOv8
- PaddleOCR
- InsightFace
- Liveness detection
Core project technologies named in the CV. Hosting topology, deployment framework and production capacity are not specified here.