Deepn
A couples coaching app built by three AI tools in a row. I audited it, cut it down to the code actually in use, and rebuilt it as one clean, organized codebase.
- Client
- Deepn
- My role
- Engineer brought in to rescue the codebase
- When
- June 2026 to present
- Status
- Ongoing
The problem
Deepn is a growth and communication tool for couples. Its core feature is a set of AI-guided "trigger" exercises: a user describes a situation and gets coaching on how to handle it better.
The product had been built by three AI tools, one after another: first Replit, then Codex, then Claude Code. Each tool added to what the last one left behind. By the time I joined, the repository was large and incoherent, and it was hard to tell which code was even being used.
This is a common place to end up. AI tools are very good at getting a product to a working demo. They are much less good at keeping the structure clean as features pile up, especially when a different tool picks up where another stopped.
What I did
- Audited the whole repository and stripped it down to the code that was actually in use.
- Restructured it into a monorepo. The marketing site moved out into Astro, the app became its own package, and the server was cleaned up.
- Rewrote the database schema with proper relations, then rebuilt most features on top of it.
- Optimized queries across the app.
The result
The product now has a structure a developer can understand and change safely: a separate marketing site, app and server, on a schema that reflects how the data actually relates.
