A financial platform moved from a legacy monolith to a modern microservices architecture — with AI agents in the workflow and an engineer behind every decision.
The company's original system had outgrown its architecture as the business prepared for its next stage of growth. The platform had to become faster, more scalable, and ready to handle a lot more traffic. A compact team had to deliver that while the legacy system kept running. We put AI agents into the engineering workflow, from repository analysis to code review and documentation. Engineers stayed accountable for every change.
Type
AI-Assisted Software Development / Legacy Migration
Platform
Web
Industry
Fintech
A fintech company building a platform for institutional digital asset trading and payments. Its first version was an MVP: a monolithic system built mostly in-house, with new functionality added whenever the product needed to move faster. The MVP worked well enough to launch and support the product in production.
Then came investment, and with it a new stage of growth. The company planned to expand into new markets, add services, and bring on more partners. For a trading platform, that kind of growth puts real pressure on the system: more customers trading at once means the infrastructure has to keep up without slowing down.
The monolith wasn't built for that. New functionality required faster iteration, cleaner service boundaries, and infrastructure that could scale independently — three things the existing architecture couldn't deliver. Some parts of the codebase had also been built quickly, including with AI, so they needed a closer review before the team could build on top of them.
Together with the client, we decided to migrate to a microservices architecture.
A compact team handled a migration that would normally require a much larger one. Across the team, 85 to 100% of pull requests were AI-assisted, and code review followed the same pattern — AI agents took over the first-pass review entirely on the architecture workstream, up from about half at the start of the project.
Analysis progressed the fastest. Agents audited the existing services and built a working map of the system in under an hour — work that used to take weeks of reading repositories by hand. According to the solution architect, unit test writing time dropped by roughly 90%, from about four hours to 15–20 minutes for a small task. On the front-end workstream, task lead time dropped from about a day and a half to half a day.
Code review got stronger, too. AI agents flagged issues before a human reviewer opened the pull request, while a security agent in the CI/CD pipeline checked every change for leaked secrets. Architecture documentation, diagrams, and ADRs stayed up to date instead of falling behind.
On one workstream, an architect and a single developer delivered the volume of work that would normally take five or six developers.
Intellectsoft used Claude across the engineering workflow, from the initial architecture audit to documentation, making it central to the team's development process. The team included a solution architect, four developers, a full-stack developer, a business analyst, and a DevOps engineer. AI adoption started with chat assistance and moved to agentic development within weeks, with Claude Code becoming the team's primary coding tool.
Architecture and analysis. Before designing each new service, the architect pointed an agent at every repository in the system. Each collected findings against a checklist, and a separate agent merged the results into one analysis. The team mapped the legacy system in about an hour. The architect then reviewed the findings by hand.
Code. Developers used Claude Code for agentic development, and most pull requests across the team were AI-assisted. The team had one rule: if you don't understand what the agent wrote, it doesn't go in. To keep pull requests reviewable, the team split tasks into smaller pieces.
Workflow automation. Through MCP servers, agents connected to Jira, GitHub, Slack, and the databases. They created and closed tasks, wrote task descriptions, and opened pull requests with a Slack message that tagged the reviewer. Each automation saved a little time, adding up to roughly half an hour per person per day.
Review and security. Every engineer configured agents differently, but the team shared two: one for code review and one for security review, alongside a coding agent built around a skill for the team's programming language. The architect ran a larger set on top of that — including agents for system design, documentation, and requirements analysis — used to draft user stories and use cases. Custom agents reviewed every pull request and ranked findings from blocker to low priority. Frontend and backend changes each passed a dedicated reviewer, and a security agent checked every change. Another security agent in the CI/CD pipeline looked for leaked secrets. Engineers made the final decision.
Testing. Agents wrote unit tests and helped draft test cases and test plans for integration and end-to-end testing.
Documentation. Agents produced Mermaid diagrams, drafted ADRs from meeting notes, and updated architecture documents when a task closed. The business analyst also maintained a Markdown documentation repository with a dependency graph that agents used as a project knowledge base. When a new feature conflicted with an earlier decision, they flagged it.
Process. The team built a shared AI hub and tracked how each person used AI: pull requests, reviews, tests, defects, and which model handled which task. Heavier models went to architecture work, lighter ones to routine documentation — Claude Opus for complex design decisions, lighter Claude models for routine writing.
Go
React
Microservices
Claude Code
PostgreSQL
Kubernetes
Eurostar
By offering innovative automated solutions, we've helped this high-speed railway leader significantly improve passenger information processing, enhance onboard amenities, and ensure timely...
By offering innovative automated solutions, we've helped this high-speed railway leader significantly improve...
view more
AF Gruppen
One of the biggest construction companies in Norway gets a digital transformation of their ERP with three spot-on solutions.
One of the biggest construction companies in Norway gets a digital transformation of their ERP with three spot-on...
view more
Ernst & Young
Our application for E&Y elevates industry expertise to a new level of thought leadership. The 'Forecasts in Focus' app empowers users with insightful data, enabling more informed investment decisions...
Our application for E&Y elevates industry expertise to a new level of thought leadership. The 'Forecasts in Focus' app...
view more
Smart Luxury Hotel
An innovative, award-winning Asian hotel chain introduces a unified smart room application to streamline guest services across all its luxurious suites.
An innovative, award-winning Asian hotel chain introduces a unified smart room application to streamline guest...
view more
Airspace Spare Parts Marketplace
Revolutionary tracking, management, and procurement platform for civil airlines spare parts that utilizes physical assets tokenization on blockchain.
Revolutionary tracking, management, and procurement platform for civil airlines spare parts that utilizes physical...
view more
Jaguar Land Rover
The JRL Talk solution is a multilingual mobile application that lets the company’s representative leap over language barriers in a matter of seconds, ensuring clear and productive conversations.
The JRL Talk solution is a multilingual mobile application that lets the company’s representative leap over language...
view more
