Jun 2025 – Present · 1 yr · On-site

• · Technical lead for Smart QC — an AI/ML document verification platform built to intercept fraudulent Aadhaar enrolments. Shipped Phase 1 to production: live Birth Certificate, PAN, and Driving License verification, detecting 5% fraudulent documents in early production data.
• · Designed a fully async, multi-service orchestration pipeline — Smart QC Service → Model Pipeline → Packet Document Service → Document Integration Service — with Kafka as the messaging backbone and Couchbase for document deduplication. Collaborated directly with the Couchbase engineering team to integrate vector search (pending rollout).
• · Architected the Packet Document Service: a centralised document caching and encryption layer serving all enrolment front-ends, legacy pipelines, and Smart QC — handling HSM-backed encrypted-at-rest storage, decryption on retrieval, and legacy doc-ID resolution at platform scale.
• · Built canary-mode model execution with result versioning — multiple ML model versions run in production simultaneously, enabling data-driven promotion decisions without service downtime or rollback risk.
• · Eliminated a critical distributed deadlock across 15+ microservices caused by a Kafka 3.3 transaction bug. Executed a full cluster migration to Kafka 4.0 overnight — halting gateway inflow, migrating, and gradually reopening traffic — restoring full pipeline stability with zero data loss. Recognised as "Bug Hunter" by UIDAI leadership.
• · Drove OpenBao Vault adoption from PoC to pre-production: architected dynamic secret rotation and Vault Agent Sidecars for Kubernetes-native secret injection. Architected PKCS#11 HSM-backed Auto-Unseal and integrated HSMs as the encryption/decryption boundary for all PII data — meeting UIDAI's highest security compliance tier.
• · Lead a 7-member backend team: structured code reviews, architecture discussions, API contract standards, and cross-functional coordination with a dedicated AI/ML team.

Jun 2023 – Jun 2025 · 2 yrs · On-site

• · Engineered 99.9% message delivery reliability at national scale — built a Kafka producer-per-thread architecture with custom Dead Letter Topic (DLT) recovery, absorbing enrolment burst loads without packet loss across the Aadhaar pipeline.
• · Cut database latency 15–20% in production: refactored the persistence layer with Redis caching, HikariCP connection pool tuning, and JPA sharding — eliminating recurring connection-exhaustion incidents under peak load.
• · Designed a Replay-Inflight mechanism for zero-risk failed-packet reprocessing — safely re-drives failed multi-stage packets across microservices with strict deduplication guarantees, reducing reprocessing error rate to zero.
• · Replaced REST with gRPC for all east-west inter-service communication and introduced GraphQL for client-facing APIs — reducing inter-service overhead and decoupling data contracts from service internals.
• · Built fault-tolerant consensus nodes using Apache Ratis (RAFT), enabling state-machine replication across distributed nodes and eliminating single points of failure in critical enrolment services.
• · Prevented duplicate UID enrolments at scale by implementing Solr-powered demographic deduplication with complex multi-field matching logic, suppressing invalid UIDs before they entered the identity registry.
• · Ensured distributed transaction atomicity via Chained Kafka-Database Transactions — preventing partial state writes across service boundaries and reducing rollback incidents to zero.
• Built real-time Grafana dashboards and ELK pipeline audit trails for production observability; shipped microservice updates via Jenkins + ArgoCD GitOps on Kubernetes with full rollback capability.