The Codebase AI Couldn't Read

The Challenge

Lemonsoft team started using the AI tools in coding back in 2025. They tested out Copilot, ChatGPT, Cursor and understood the productivity boost that these tools brought to their development: routine work had decreased, and team sentiment was positive. On paper, everything looked good.

But something was quietly wrong, and Janne Tammi, CTO, couldn't name it for a long time. The problem wasn't the tools themselves; it was what they didn't know. Each AI session started from zero, with no memory of the architectural decisions made the week before and no recollection that a certain approach had already been tried and abandoned a month ago. With ten developers each starting fresh ten times a day, that added up to a hundred zero-start sessions every day. The watching and correcting got old fast, and trust eroded slowly.

There was also a deeper problem: the AI couldn't understand Lemonsoft's mature, complex codebase at all. Lemonsoft's software has been built and refined over years — millions of lines of business logic, architectural patterns, and accumulated decisions that reflect how the company actually operates. That knowledge doesn't live in any single file or document. It lives in the codebase itself, and in the heads of the people who built it. When an AI tool generated code, it produced something syntactically correct but architecturally foreign — code that didn't fit the patterns, didn't respect the conventions, and didn't account for the decisions that had already been made. Developers would catch it in review, correct it, explain the context to no one, and start over the next day.

This created a quiet but persistent bottleneck. Senior developers became the connective tissue between AI output and architectural reality, fielding corrections and questions that the tools should have been able to answer on their own. The more AI tools were used, the more human time was consumed keeping the output aligned with a system the tools couldn't see. The productivity gains were real, but they were smaller than they should have been and the gap between what AI promised and what it delivered in practice was growing harder to ignore.

Why ModernPath

In 2025, Janne spoke with ModernPath and walked through the situation together with the company founder and AI-development pioneer Pasi Vuorio. The conversation wasn't about which tools to use or how to prompt them better. It was about the underlying problem: a codebase that AI tools couldn't read, and an organization paying the cost of that gap every single day. True to his style, Pasi was promised he would solve the problem. He would get Lemonsoft's mature codebase into a form that AI could understand and work with reliably.

"I wish I had shared his confidence," Tammi says. "Honestly, I didn't believe the project would succeed."

The skepticism was reasonable. Lemonsoft's codebase isn't a greenfield project or a clean microservices architecture. It's the kind of system that carries years of real-world decisions, workarounds, and domain-specific logic that no documentation fully captures. Making that legible to AI — not just at the surface level but in terms of structure, patterns, and intent — is a different challenge from anything a tool license or a prompt template can solve. The codebase itself needed to become a structured knowledge layer, something AI agents could query accurately and build on reliably. That was the step Lemonsoft had to take before anything else could change, and it required trusting that someone else could do what hadn't been done before