Maintainable and secure Vibe coding
Project description
"How can software development teams effectively maintain and extend "vibe coded" PoC codebases to ensure efficient collaboration, scalability, and long-term sustainable development?" - Innitial question proposed by our client
Context
Modern software development increasingly begins with rapid prototyping—proof-of-concepts (PoCs) built quickly to validate ideas and secure stakeholder buy-in. These "vibe coded" codebases prioritize speed over structure, often resulting in minimal documentation, inconsistent patterns, and technical debt. While this approach accelerates initial delivery, it creates significant challenges when transitioning to production-ready systems.
Development teams frequently inherit these PoCs and face a critical dilemma: rebuild from scratch or refactor. Rebuilding risks timeline delays and scope creep, while refactoring demands careful strategy to avoid worsening existing issues. Without proper architectural clarity or coding standards, teams struggle with maintaining and dealing with such projects.
The client presented this challenge without a specific solution in mind. Our team had to independently research, analyze, and develop possible solutions and approaches for tackling this problem.
Results
On top of the research, our project delivered a AI-powered code review system built entirely from scratch to address the challenges of maintaining "vibe coded" PoC codebases. The key outcome is a flexible, automated PR reviewer that analyzes code quality, identifies issues, and provides actionable feedback to development teams transitioning from prototyping to sustainable development.
The system supports both local models through Ollama and hosted ones like OpenAI and Gemini, making it adaptable for teams with varying infrastructure requirements, budget constraints, and data privacy needs. This dual-mode architecture ensures accessibility across different needs.
Beyond analysis, the tool generates concrete code fixes for identified issues, accelerating the refactoring process and reducing manual intervention. This capability transforms the reviewer from a diagnostic tool into an active development assistant, directly addressing technical debt while preserving developer control through reviewable suggestions.
The solution aligns with TRL 4: demonstrating validated functionality in laboratory environments with clear pathways toward production integration. Our solution emphasize that sophisticated automated intelligent tooling can minimize the gap between PoC velocity and developers struggle.
About the project group
This was an 18-week project. Our group consists entirely of software students with backgrounds in either AI or game design. We had to work both individually and as a team depending on the stage of the project, so our previous experience in such environments was a huge advantage. This, combined with the agile way of working, helped us be flexible and adaptable to the situation.