Sub-Second BNPL Microservices Migration on AWS

The company’s flagship product is a “split-in-four” checkout API and merchant portal that lets shoppers break purchases into interest-free installments while retailers get paid up front. Early traction was strong, but the platform’s original monolith — an overgrown ASP.NET MVC app — was straining under peak-season traffic, reconciliation delays, and an ever-growing list of compliance stories. When the CTO came to us, he needed two things before the next investor demo: sub-second loan decisions and more predictable AWS costs.

We proposed a light-touch rebuild, rather than a full rip-and-replace rebuild. Over a six-week discovery sprint, we carved the credit-decision engine out of the monolith, rewrote it in ASP.NET Core as a stateless API, and fronted it with an event queue, so approvals returned in under 300 ms even during flash sales. Parallel teams decomposed reporting jobs into Lambdas and shifted static merchant assets to S3+CloudFront, trimming nightly batch windows and halving storage spend. The result: a leaner, auditable BNPL backbone ready for scale without a single minute of service blackout.

• Solution Architect (credit-decisioning & AWS microservices)
• 2 Backend Engineers (C# / ASP.NET Core)
• DevOps Engineer (IaC, observability, cost optimisation)
• QA Automation Engineer
• Data Engineer (reporting & reconciliation pipelines)
• Project Manager

• Monolithic bottleneck. A decade-old ASP.NET MVC codebase handled checkout, underwriting, reporting, and email in one IIS pool; CPU spikes during flash sales pushed loan decisions past two seconds and triggered cart abandonment.
• Latency-sensitive credit rules. Fraud scoring relied on synchronous calls to three external bureaus, timeouts stacked, and the retry logic blocked the entire request thread, throttling peak throughput.
• Overnight reconciliation drag. A single 90-minute SQL job merged daily authorisations, refunds, and chargebacks; any hiccup delayed settlement files to the bank and set off manual triage at 6 a.m.
• AWS cost creep. Orphaned EBS snapshots, over-provisioned RDS replicas, and idle ECS tasks added ~30% to monthly spend with no performance benefit.
• Compliance pressure. New CFPB guidance required field-level audit trails for every credit decision, but the legacy schema stored only a final status flag — no intermediate score or rule hit.
• Zero-downtime mandate. Merchants depended on the API for point-of-sale funding; even a five-minute maintenance window risked lost checkout revenue and SLA penalties.

• ASP.NET Core microservices on AWS ECS Fargate
• Amazon SQS event queues for async credit decisions
• PostgreSQL (Amazon Aurora) plus DynamoDB side store
• AWS Lambda functions for reconciliation and reporting jobs
• S3 + CloudFront for static merchant assets
• Terraform IaC, GitHub Actions CI/CD pipeline
• CloudWatch Logs, Metrics & X-Ray tracing for observability

The Devox Software team slid a service layer around the legacy checkout flow and carved its slowest organs out one at a time.

• Split the heartbeat. We lifted credit-decision logic into a stateless ASP.NET Core service, fronted it with SQS. The checkout service now posts a message and moves on, while the new engine pulls bureau data in parallel, scores the request, and replies in < 300 ms — even on Black-Friday traffic.
• Make data audit-proof. Every rule hit, bureau score, and final verdict streams into Aurora with an immutable JSON snapshot; DynamoDB holds the hot session token so the UI can flash an approval before the card reader cools.
• Rebuild reconciliation for daylight. Nightly SQL blocks became three Lambda steps — authorise, settle, chargeback — with each drop-file signed and stored in S3. The 90-minute job now clears in under nine, and finance walks in to a green dashboard.
• Slim the cloud bill. Terraform tagged idle EBS, right-sized RDS replicas, and capped ECS scale-out. Monthly AWS spend fell a third while p95 latency fell sixfold.
• Deploy without a hiccup. GitHub Actions push a blue container set, Fargate flips the target group, and X-Ray traces confirm clean hops before we drain the old task — merchants never notice the hand-off. In four sprints, the BNPL engine went from monolith fatigue to microservice muscle, ready for the next investor demo and the holiday traffic that follows.

“What began as a request to speed up the backend quickly evolved into a strategic collaboration. We spotted broader opportunities — from architecture gaps to AWS cost inefficiencies — and proposed a full modernization of critical services without a single minute of downtime. That kind of proactivity is what earned us the client’s trust.”

— Vadym Zotov, Commercial Director, Devox Software

BUSINESS OUTCOMES

• Checkout flows feel instant. Shoppers now see “Approved” while their items are still in the cart; merchants report fewer abandoned orders during peak promos.
• The finance team starts the day with everything in the green. Reconciliation files arrive in their inbox before 6 a.m.; no more scrambling to explain missing settlements.
• Investor narrative sharpened. Demo dashboards show real-time decision speeds and declining costs — key proof points for the upcoming funding round.
• Support tickets drop. Fewer “payment pending” calls and zero “API timeout” complaints since the switchover, freeing the ops team for strategic work.

TECHNICAL OUTCOMES

• Decision latency cut from seconds to sub-300 ms through SQS buffering and parallel bureau calls.
• Nightly processing window shrank 90%. Lambda pipeline clears settlements in under ten minutes instead of an hour-plus batch.
• AWS bill trimmed by roughly one-third via right-sizing, snapshot cleanup, and auto-scaling only when queues surge.
• Audit-ready data lineage. Every credit rule, bureau hit, and response is stored immutably with a JSON snapshot and SHA hash.
• Zero downtime across five blue-green releases, validated by X-Ray trace health before traffic flips.