CVXpert

AI tools that rewrite CVs tend to make things up, like a metric nobody measured or a tool the candidate never used. I had the idea for CVXpert in 2025 and wrote a first plan. In October 2026 I came back to it and built it: a CV reviewer that cites its sources and won't invent facts.

What I built

I built CVXpert alone, as an LLM workflow. It parses the CV, reviews it against 24 guidelines drawn from university career services, can tailor it to a job description, and rewrites it. I kept it a simple workflow instead of a multi-agent system, because the task is a fixed sequence with one optional branch. Every finding names the guideline it rests on.

Checks in code guard the rewrite. If the rewrite contains a number the CV doesn't, the model gets one retry, and the run fails if the number is still there. Names and tools the CV doesn't mention are listed for the user to check. The user can also reject findings and confirm which job-description gaps are really true of them, and only confirmed gaps may enter the rewrite. The tests run the real workflow on fake models, so CI never calls a paid model.