How Workstation’s AI Agents and Edge Gateway Took DIY Tax Return from Dev to Production

DIY Tax Return is now in production. UK taxpayers can prepare and file their Self Assessment with AI help at tax.diytaxreturn.co.uk. Our technology partner Workstation worked with us from the first commit to launch day. This post explains how their AI development, QA and deployment agents, and the WSL WAF / API Gateway / Edge Gateway, took the product through every environment from Dev to Prod, checked against HMRC’s Making Tax Digital (MTD) requirements at each stage.

A technology partner from inception to launch

Workstation built the delivery platform behind DIY Tax Return: the Kubernetes clusters, CI/CD pipelines, observability stack and edge gateway. Their AI agents worked alongside our engineers throughout. That meant a small team could build, test and harden an HMRC-connected product at a fraction of the usual cost and time.

AI agents across the whole lifecycle

Development agents

AI coding agents wrote and refactored code under engineer review. That work covered the FastAPI backend, the Next.js web app, the Flutter iOS app, and the HMRC MTD integration. Every pull request also went through an AI reviewer that checks correctness, security and HMRC-specific rules before merge.

QA agents

QA is where AI saved the most time. Automated Playwright suites run the real user journey from start to finish:

• Registration and login, including one-time-passcode flows
• Bank statement upload, AI extraction and verification
• AI transaction classification and reconciliation
• Tax calculation and the HMRC filing journey
• Role-based admin access and billing flows

These suites run on every change and again every night. Results are published to a live QA dashboard. When a build breaks, an AI build-fixer agent diagnoses the failure and proposes a fix, so engineers spend their time reviewing rather than debugging.

Deployment agents

Releases move through five isolated rings: dev, int, test, acc and prod. A ring promoter only moves a build forward when the ring before it is green. Every environment is declared in Helm and GitOps, so production is a reproducible copy of what was tested, not something configured by hand.

WSL WAF, API Gateway & Edge Gateway

All traffic to DIY Tax Return, in every environment, passes through Workstation’s WSL edge gateway. It gives us:

• Web Application Firewall: filters malicious traffic before it reaches the app
• API gateway routing: weighted backends per environment, so int, test, acc and prod stay isolated
• TLS termination and edge caching: fast, secure delivery of the web app and this marketing site
• Declarative configuration: gateway rules live in version control, are checked in CI and are scanned for drift automatically

Built against HMRC’s MTD requirements at every stage

Software that files Self Assessment through HMRC’s Making Tax Digital APIs has to meet HMRC’s technical and fraud-prevention requirements. We made those checks part of the pipeline:

• Fraud-prevention headers validated automatically: every change to our Gov-Client-* / Gov-Vendor-* headers (web and iOS) is tested against HMRC’s official validator, and a weekly scheduled run catches rule changes from HMRC
• HMRC sandbox end to end: the full filing journey is tested against HMRC’s sandbox before any build can be promoted
• Secure HMRC connection: OAuth 2.0 authorisation with HMRC, with the user’s National Insurance number checked before connecting
• You review before you file: you check your figures before anything is sent to HMRC, and every submission is audit-logged
• Resilience: automated database backups and disaster-recovery restore drills

The aim was never just to “go live”. It was to reach production with the same checks passing in every ring, so the first real filing behaves exactly like the thousandth.

Built to scale, and to keep costs down

DIY Tax Return is built to handle thousands of UK taxpayers filing their Self Assessment with AI, including the rush before the 31 January deadline. Using AI agents for development, QA and operations, and generative AI for the product itself, kept costs down in several ways:

• Smaller team, faster releases: agents do the repetitive build, test and deploy work
• Lower cost per user: efficient Kubernetes infrastructure and edge caching instead of expensive managed platforms
• AI that you can monitor: every generative AI call is traced and measured, so model costs and quality stay under control

Those savings go straight to our customers as one of the most affordable ways to file a UK tax return.

Technology partner: Workstation — AI development, QA & deployment agents and WSL Edge Gateway

← Back to all articles