MIRKO

Replacing a Core Payments Engine With Zero Customer-Visible Downtime

A UK retail bank replaced its core payments engine serving 4.1 million accounts. We delivered the new .NET platform on Kubernetes, a parallel-run reconciliation layer and an LLM-based assistant for operations teams, reaching 99.99% availability.

Industry
FinTech
Country
UK Retail Bank
Team Size
22
Services
Custom Software Development, QA & Testing, AI Integration
Technologies
.NET, Kubernetes, PostgreSQL, Kafka, Azure, OpenAI
Replacing a Core Payments Engine With Zero Customer-Visible Downtime

Overview

About The Project

The bank's payments engine dated from the early 2000s. It processed faster payments, direct debits and card settlements for 4.1 million accounts, and every change risked an outage visible to customers and regulators.

Over eighteen months we built the replacement on .NET and Kubernetes with PostgreSQL and Kafka, ran both engines in parallel with automated reconciliation, and migrated traffic scheme by scheme. An operations assistant built on an LLM helps the support team investigate payment exceptions with full audit trails.

The Challenge

Change A RunningEngine Without Stopping It

Regulators require continuous availability and complete traceability. The legacy engine had undocumented edge cases accumulated over twenty years, and any discrepancy between old and new behaviour had to be explained.

Operations teams investigated exceptions with manual database queries, which did not scale to the parallel-run period where every difference needed a root cause.

Change A Running Engine Without Stopping It

Our Solution

Parallel Run,Reconciliation And An Ops Copilot

Every payment flowed through both engines for months. A reconciliation service compared outcomes field by field and classified differences; the ones caused by legacy quirks became documented rules, the rest became fixes. Traffic moved scheme by scheme behind feature flags with instant rollback.

For the operations team we integrated an LLM assistant into their case tool. It reads a payment exception, pulls the relevant ledger entries and scheme rules, and drafts an investigation summary with citations. Every suggestion is reviewed by an analyst, and the assistant is evaluated weekly against closed cases.

Parallel Run, Reconciliation And An Ops Copilot

Key Features

What WeDelivered

The Impact

Results ThatMatter

Customers never noticed; the regulator received the evidence.

99.99%

Availability

0

Customer-visible downtime

4.1M

Accounts migrated

-45%

Exception investigation time