FinTech · Banking: First Abu Dhabi Bank (UAE)
Cheque Clearance Automation for First Abu Dhabi Bank
Event-driven cheque clearance on Kafka / Azure Event Hubs, covering preprocessing, YOLO detection, OCR and signature matching through microservice consumers.
- Role
- AI Engineer - Cloud & Platform Lane
- Owned pipeline and platform: Kafka event backbone, deterministic rules engine, audit chain, signature embedding serving, Azure landing zone, MLOps and disaster recovery.
- Industry
- Banking: First Abu Dhabi Bank (UAE)
- Timeline
- 12 Weeks (Phase 1) - phased roadmap to 8 months
Business impact
- target throughput
- 20,000/day
- languages
- Arabic + English
- corpus
- 3,000-5,000 cheques
- audit trail
- Tamper evident
- rollout
- 5% - 25% - 100%
Overview
Event-driven cheque clearance platform for First Abu Dhabi Bank built on Kafka / Azure Event Hubs, where partitioned topics stream cheques through a chain of microservice consumers: image preprocessing, YOLO detection, OCR extraction, signature matching, then automated clearance.
Topic partitioning and consumer groups were tuned to remove stalling under load. Deployed on AKS with Helm, private networking and autoscaling for secure, bank-grade throughput.
Technical architecture
- Ingress
- Application Gateway with WAF, API Management AI hub gateway, Entra ID managed identities
- Event Backbone
- In VNet Kafka with private mTLS, idempotent consumers and dead letter queue
- Orchestration
- Deterministic state machine on AKS calling checks and the LLM as tools
- Vision & Extraction
- Image QA, MICR reader, OCR/ICR and vision LLM structured extraction
- Risk & Verification
- In cluster signature ML, tampering forensics, sanctions screening, beneficiary matching
- Data & Storage
- Blob images, PostgreSQL case state, Redis idempotency keys, immutable WORM ledger
- MLOps
- Model and prompt registries, golden dataset eval harness with regression gate, drift monitoring
- Resilience
- In country DR with geo replicated storage and database, documented RTO/RPO
Key features
- Multi modal extraction: OpenCV preprocessing, YOLO region detection, MICR codeline reading and VLM structured JSON output
- CAR/LAR reconciliation - courtesy amount cross checked against legal amount in Arabic and English
- Signature verification at scale using ONNX embeddings with 1:N matching over pgvector
- Fraud and tampering forensics with calibrated FAR/FRR thresholds
- Beneficiary name matching using fuzzy and embedding techniques across Arabic and English
- Deterministic rules engine owning every final decision with a governed reason code taxonomy
- Confidence gate routing low confidence and complex mandates to human review
- Tamper evident SHA 256 audit chain and WORM evidence pack per cheque
- Mandate reasoning - Arabic and English mandate letters parsed into structured signer groups and limits
Challenges solved
- Arabic handwriting accuracy - measured early against a golden set with a swappable vision model gateway
- Keeping AI out of the decision path - LLM output is structured and advisory only, the rules engine decides
- Signature threshold calibration delivered as a risk committee memo with ROC curves, not hard coded numbers
- Sizing for a peaked morning arrival profile using KEDA autoscaling tuned on Kafka consumer lag
- Central bank retention satisfied with a WORM audit store and SHA 256 chaining
Tech stack
- Python
- Azure Kubernetes Service
- Apache Kafka
- PostgreSQL + pgvector
- Azure OpenAI (GPT 4o Vision)
- Azure AI Document Intelligence
- Azure AI Search
- ONNX
- Computer Vision (YOLO, OpenCV)
- Tesseract OCR
- Azure Key Vault (CMK)
- Azure API Management
- KEDA Autoscaling
- Docker
- Azure DevOps