The AI Lending Platform Transformation

In early 2024, AI Lending Platform — India's fastest-rising AI lending platform — was staring at disaster weeks before launch. “We expect 4,000+ concurrent users and thousands of real-time AI credit decisions per minute,” the CTO said. “Our on-premise MySQL is already dying at 500 ms latency. One outage and we lose customers, trust, and RBI compliance.”

• Legacy Windows servers and bare-metal MySQL with zero failover.
• Queries routinely exceeding 500 ms — unacceptable when conversion drops 1% per 100 ms.
• No encryption-at-rest strategy for borrower PII.
• Windows-centric enterprise teams with deep Active Directory, SQL Server, and PowerShell dependencies.
• Zero chance of surviving launch traffic on the existing stack.

We delivered a completely cloud-native, RBI-compliant platform that feels 100% native to Microsoft enterprise environments while delivering open-source performance and cost.

• Primary production on Amazon RDS for PostgreSQL Multi-AZ, chosen for cost and performance.
• Identical architecture pre-validated and ready for Amazon RDS for SQL Server or SQL Server Always-On on EC2 Windows.
• 200 tables, 140 constraints, 60 indexes — fully compatible with SQL Server data types and T-SQL patterns.
• p95 latency dropped from 500+ ms to under 20 ms at 4,000 concurrent connections.

60% of the ECS cluster runs Windows Server 2022 containers on an optimized AMI. Critical .NET Framework 4.8 and .NET 6/8 microservices — loan scoring and the compliance rules engine — run natively on Windows nodes, with zero code changes for AI Lending Platform's legacy Windows services; they were simply containerized and deployed.

• On-premises Active Directory federated via AWS Managed Microsoft AD + IAM Identity Center (AWS SSO).
• All developers, DevOps, and finance teams log in to the AWS Console, CLI, and internal tools using their existing AD credentials.
• Group-based permissions (e.g. “LoanOps-Admin”, “Compliance-Audit”) mapped directly to IAM policies.

• A shadow Amazon RDS for SQL Server instance is kept in sync via DMS CDC for downstream Windows-only reporting tools (SSRS, Power BI DirectQuery, legacy .NET apps).
• All encryption, HA, and backup patterns are identical to the production PostgreSQL setup.
• AI Lending Platform can flip any new module to SQL Server tomorrow without re-architecting.

• The entire IaC secondary layer is written in PowerShell DSC + AWS Tools for PowerShell, alongside Terraform.
• GitHub Actions runners (Windows-latest) execute PowerShell/.NET scripts for secret rotation, compliance report generation, Windows patch compliance validation, and automated failover testing.
• Zero training required — AI Lending Platform's Windows admins became cloud admins overnight.

• Patch Manager automatically patches Windows containers and EC2 Windows instances.
• Session Manager replaces RDP/SSH with audited, bastion-less access.
• Inventory + State Manager enforce golden Windows AMIs across the fleet.
• Run Command executes PowerShell remediation scripts in under 30 seconds, fleet-wide.

Amazon EBS gp3 volumes, KMS-encrypted, are attached to Windows ECS tasks for high-speed local SSD caching of KYC documents, Windows event logs and application tracing, and temporary .NET memory dumps that are auto-encrypted.

A mixed Linux/Windows ECS cluster auto-scales to 4,500+ connections. Windows services use ElastiCache for Redis natively via StackExchange.Redis, while CloudFront and Route 53 deliver sub-second responses India-wide.

• AWS KMS + Secrets Manager with 90-day rotation via PowerShell Lambda.
• All Windows instances domain-joined to AWS Managed Microsoft AD.
• CloudTrail + GuardDuty + quarterly IAM/AD access reviews.
• Hardware MFA and privileged access logging for all root and AD admin actions.

• AWS DMS + SCT converted MySQL to PostgreSQL.
• A parallel DMS CDC pipeline kept the SQL Server shadow instance in sync.
• Blue/green Windows + Linux ECS deployments completed total cutover in under 12 minutes.
• Tested fallback to on-premise Windows/SQL Server in read-only mode.

• 4,300+ concurrent users at peak.
• Windows-based .NET scoring services processed 180,000+ AI decisions per hour.
• p95 query latency: 18 ms.
• End-to-end offer generation: under 90 ms.
• 99.999% uptime — one brief Multi-AZ failover lasting 6 seconds, with zero customer impact.

• 50% TCO reduction vs on-premise Windows + SQL Server licensing.
• Windows patch compliance went from 63% to 100% automatically.
• PowerShell-driven monthly cost & security reports now take 4 minutes instead of 4 days.
• AI Lending Platform's Windows team now deploys production changes faster than their Linux team.