Skip to content
jCode

Fix or rewrite an application built with AI?

Start with an audit: targeted repairs if the foundations hold, gradual replacement if only some parts are fragile, a rewrite as a last resort. This guide draws on 10 years of web development, application takeovers included, and daily work with coding agents.

Published 13 October 2026 · Jean-Baptiste Thery · EURL jCode

001
Step 1

AI-generated code tends to fail in the same places. When a researcher scanned 1,645 apps built with Lovable in 2025, about 10% let anyone read or write their database tables, because row-level access rules were missing.

What the audit checks

Authorisation and database access rules, secrets in the code and its history, dependencies, the separation between development and production, then tests and the deployment process.

Why secrets come early

In GitGuardian's March 2026 report, public commits co-written with Claude Code exposed secrets in 3.2% of cases, more than twice the 1.5% average.

What you receive

A written report of the issues, their risk and the work needed, estimated by work package. When I take over an application, this fixed-price audit is the first step.

Back to the steps

002
Path A

Repairs fit when the foundations hold and the problems are local.

The signs

The data model fits the business, access rules can be added without a redesign, and most defects sit in a few places.

What it looks like

Close the security gaps first, then fix defects by priority and add tests around each part you change. The application keeps running throughout.

Back to the steps

003
Path B

Gradual replacement fits when some parts are sound and others are not.

The signs

The application has users and data to keep, but some modules are too fragile to repair. A stable interface can sit between the old and the new parts.

What it looks like

Replace one module at a time behind that interface, migrate its data, then remove the old code once the new part runs in production.

Back to the steps

004
Path C

A rewrite is a last resort, not a starting point. Sometimes it is still the cheapest path.

The signs

The data model or the security foundation is wrong throughout, nobody can explain how the code works, and every fix breaks something else.

What it looks like

Keep the product knowledge, the data and what users rely on, then rebuild in small releases, with tests and a clear architecture from the first one.

Back to the steps

005
With AI

Coding agents help with the repair too, under the same rules as any other change.

How I use them

They help me map the codebase, write the missing tests and draft fixes on the points the audit raised. I review and test every change before delivery.

Back to the steps

007
Contact

Let's talk about your project.

Tell me what you want to build or improve, your budget and your target date.

  • Fully remote
  • English or French
  • France · UK · US