My team went too fast with AI coding, and now we're drowning in tech debt
About six months ago, I was excited about shipping features quickly using AI-assisted development. We moved fast, iterated constantly, and hit milestones I would've called impossible a year earlier.
Then I tried onboarding a new developer last month. The codebase was a maze. Not because we lacked skill—we have good developers—but because the AI made it too easy to add layers without cleaning up what was underneath.
We've now started a conversation about rewriting our core modules the old way: slower, more deliberate, with stricter code review. It's humbling to admit, but sometimes fast feels expensive in retrospect.
This got me thinking about whether there's actually a market for tools that help small teams manage code quality when everyone's using AI. Something lighter than enterprise solutions, that actually fits a micro SaaS or one-person company workflow.
Has anyone tried building (or buying) something in this space? Curious if the pain point is widespread enough to be a real opportunity, or if it's just us being undisciplined.
I’m starting to see this from the solo-builder side too.
I’ve been building larger products with Codex, and the dangerous part is how cheap a new feature suddenly feels. You can add something in minutes, so the instinct is always to keep moving instead of cleaning up what is already there.
The cost often shows up later as regressions rather than obviously bad code. A change works, but it quietly affects navigation, persistent state, playback behaviour or some other part of the product that was already stable.
I think there probably is a real tool opportunity here, but I wouldn’t want another heavy “code quality dashboard.”
What I’d personally want is something that sits between the coding agent and the codebase and keeps track of:
• behaviours that must not break
• architecture decisions already made
• duplicated logic introduced by new generations
• modules that are getting too complex
• regression checks that should run after each change
• when the agent should refactor instead of adding another layer
Almost like a persistent technical memory + guardrail system for AI-assisted development.
The irony is that AI lets us accumulate technical debt faster than ever, so maybe we also need AI specifically dedicated to stopping that debt from growing.
When you looked at your codebase during onboarding, what was the biggest problem: duplicated code, inconsistent architecture, too many abstractions, or simply that nobody understood why previous AI-generated decisions had been made?