AI Code Cleanup
For product teams whose repo was written by Copilot, Cursor, or coding agents faster than anyone reviewed it. We score what to keep, what to repair, and what to delete, then repair the paths customers and compliance actually hit.
Three things happen, in this order. The order is the method.
Most of what the model produced should be deleted, not refactored. The $10,000 audit is the scored keep / repair / delete report. Load-bearing paths are identified by who depends on them, not by how much code they contain.
Customer-facing and compliance-facing paths come first. Security pass: secrets, dependencies, data leaving the building, prompt injection. Behavior pass: evals on the tasks the business actually cares about.
Cleanup without a harness recreates the mess in a quarter. The repair quote can include a Harness Engineer on a fixed milestone, and one internal owner trained on the repaired path. The audit fee does not include that seat.
One service. The scopes are slices of the same repo, sold together or cut to the part that hurts.
| Scope | What they bring you | Typical buyer |
|---|---|---|
| Application code | Dead abstractions, hallucinated APIs, duplicated modules | CTO, VP Engineering |
| Tests | Suites that assert nothing, or no tests on the path customers hit | VP Engineering |
| Agent and tool code in the repo | Checked-in agent loops, tool wrappers, and prompts with no eval | CTO |
| Secrets and dependencies | Keys in the tree, unsafe packages, generated clients that send data out | CISO, Head of IT |
A running agent that loops, or a pile of generated documents and SOPs, is a different engagement. Behavior of a live agent is a Harness Engineer or AI Ops seat. Document cleanup is not this service. A legacy product nobody generated is a rebuild, not a cleanup.
The $10,000 buys the scored keep / repair / delete report, delivered in 7 days. The report names what to keep, what to repair, and what to delete.
The repair, the harness, and the named owner are the next quote. The repair usually runs 6–8 weeks. There is no dollar figure for it on this page.
Buyer situations. These are not AIPIVT client results, and they are not the founder's historical cost-reduction record.
Payments, server-side access, and a place it runs are one repair of a generated app. The audit is the same $10,000 report.
Payments that survive a dropped webhook.
Secrets off the client. Authorization on every route.
A domain, logs, and a named owner. We do not host it.
The scored keep / repair / delete report, delivered in 7 days. It does not buy the repair, the harness, or a stay-behind engineer. Those are a separate quote. The repair usually takes 6–8 weeks. We do not publish that price.
No. We cut the system down before we touch the parts you run. A smaller codebase that passes evals is the point. A ground-up product rebuild is a different engagement.
Yes, when the tests were generated and they assert the mock, or when the path customers hit has no tests. We delete the tests that do not test and leave a suite on the path the product runs.
No. Document cleanup is out of this service. A live agent that loops is a Harness Engineer or AI Ops engagement. Book the scoping call and we will say which one it is.
A 30-minute scoping call is enough to tell whether the audit is the right first step.