My LLM Kept Making Stuff Up on Resumes. Here's How I Shut It Down.

A developer reduced hallucinations in a resume AI tool by shifting scoring from GPT-4o to deterministic Python calculations.

The model still helps with rewriting, but it now returns structured edit diffs instead of free-form claims.

Every proposed change is validated on the server before a new optimized resume record is saved.

The result is a safer pattern for LLM features: keep high-risk judgment deterministic and verify all edits before user impact.