How Marvik integrated Claude Code into Pathia’s development workflow to bring consistent, domain-aware code review across multiple repositories and technology stacks.

When AI accelerates development, review has to keep up

AI coding tools are changing how quickly engineering teams can produce software. But increasing development speed also puts pressure on another part of the process: reviewing what gets shipped.

Pathia, a digital pathology company, was already using AI within its software development process. They build software for digital pathology and their platform uses machine learning to analyze medical images, including MRIs, CT scans, and X-rays, powering AI-assisted diagnostics. 

Their engineering team could generate and iterate on code faster, but the bottleneck showed up on the other side of the workflow. Every change going to production was manually reviewed by the tech lead. 

As development accelerated, that created an imbalance between how quickly code could be produced and how quickly it could be reviewed.

The opportunity was to bring AI into that part of the software development lifecycle as well.

Marvik worked with Pathia to design and build an AI-powered PR review system using Claude Code, designed around the company’s repositories, technology stacks, and domain-specific engineering requirements.

The result was a production system deployed in approximately three weeks, providing automated and on-demand review across their existing GitHub workflow.

The challenge: scaling review across a complex software environment

Code review at Pathia involves more than identifying syntax problems or enforcing coding conventions.

The platform spans several technology stacks:

• Backend and frontend: Java, Spring Boot, Angular, and TypeScript
• Machine learning: Python and PyTorch
• Mobile: Flutter and Dart

Each repository has its own architecture and development patterns. At the same time, Pathia operates in digital pathology, where software decisions need to account for sensitive information and domain-specific data requirements.

That meant the review system needed to understand criteria such as PHI handling, tenant isolation, architecture, security, infrastructure, and patient-level data split integrity.

These were treated as project-specific review criteria and risk checks, rather than regulatory validation.

The challenge was therefore broader than automating a traditional code review. Pathia needed a system capable of applying consistent engineering and domain context across different repositories, while keeping senior engineers in control of the final decision.

Why Claude Code was a good fit

Claude Code provided a practical integration point through the official GitHub Action, allowing the review system to operate directly inside Pathia’s existing pull request workflow.

Instead of creating a separate review interface or asking developers to change tools, reviews could happen where the engineering team was already working.

The more important challenge was giving Claude enough context to understand how Pathia expected its software to be built.

Marvik used repository-level instructions and review guidelines to capture that context. These included architectural conventions, security considerations, infrastructure requirements, and domain-specific criteria that would otherwise depend heavily on the experience of individual reviewers.

Historical pull requests were reviewed together with Pathia’s tech lead to understand what useful feedback looked like in practice. Those findings became part of the review criteria used by the system.

The result was an assistant designed around Pathia’s engineering environment rather than a generic code reviewer.

How the system works: three review paths

Not every pull request needs the same level of analysis. We designed three review paths to match model capability and cost to each change.

1. Automatic review
   New PRs and updates trigger an automatic review. Smaller diffs route to Claude Haiku 4.5, while larger changes use Claude Sonnet 5. A repository-level toggle gives the team control over where automatic reviews run.
2. Standard on-demand review
   Developers can comment @sonnet to trigger a Claude Sonnet 5 review with standard reasoning effort.
3. Deep on-demand review
   For more complex changes, @claude triggers Claude Sonnet 5 with increased reasoning effort for a deeper analysis.

All three paths leave inline comments directly in GitHub. Reviews remain advisory, with developers and technical leads making the final decision.

Turning engineering knowledge into review criteria

We worked with Pathia’s tech lead and analyzed historical pull requests to capture the criteria behind their existing review process.

That context was translated into structured guidelines covering:

• ‍Architecture and maintainability: structure, patterns, and maintainability‍
• Security: sensitive data handling, authentication, and potential risks‍
• Infrastructure: configuration, resources, and environment-specific considerations‍
• Domain-specific risks: tenant isolation, PHI handling, and patient-level data split integrity

This gives Claude the engineering and domain context needed to evaluate changes against how Pathia’s software actually operates, rather than applying generic code quality rules.

From setup to production in three weeks

The first stage focused on understanding Pathia’s repositories, current pull request process, and review criteria. The team analyzed historical PRs with the tech lead and established the initial project context, review guidelines, and GitHub workflow.

The second stage introduced the automated reviewer, model routing, structured GitHub comments, and the two on-demand review paths.

The final stage focused on calibration. Reviews were tested against historical and current pull requests, compared with the tech lead’s assessment, and refined to reduce noise and false positives.

By the end of the engagement, the system was running across three repositories covering Pathia’s backend and frontend, ML pipeline, and mobile application.

Where the system stands

The AI reviewer is deployed across three Pathia repositories and has become part of the company’s software development process.

Developers can receive an automatic first review or actively request additional analysis when a change warrants it.

For the tech lead, this introduces a first layer of structured analysis before the final human decision. For the broader engineering team, it creates a consistent way to apply shared review criteria across different repositories and technology stacks.

Most importantly, the capability applies across Pathia’s software development, rather than to one isolated product feature. This was one of the key outcomes highlighted by the team after implementation.

What this project shows about AI in software development

Pathia is a useful example of how AI can enter the software development lifecycle beyond code generation.

As engineering teams produce code faster with AI, the surrounding processes also need to evolve. Review, testing, evaluation, security checks, and engineering governance become part of the same equation.

In Pathia’s case, the opportunity was code review.

The implementation also shows how quickly a focused AI system can move into an operational workflow when the integration surface is already mature and the problem is clearly defined.

“A relatively simple AI integration can create an operational change across an organization in a matter of weeks. The key is adapting the technology to how that team actually works.”‍

In approximately three weeks, Pathia went from manual review concentrated around its tech lead to an AI-assisted review layer operating across three repositories, with model routing, domain-specific criteria, and human oversight built into the workflow.

That combination matters: strong models, the right business and technical context, and an architecture designed around the way the organization operates.