Xandidate - AI candidate screening with evidence

The first version of Xandidate scored every applicant out of a hundred. Four hundred applications, ranked before lunch. Then I compared the list with a batch of CVs I had judged by hand. The scores were not wildly wrong. They were unarguable: why is this candidate a 71 and not an 80? The number had no answer, and in Europe that is barely legal.

The score went and the rubric came in. The recruiter writes what matters for the role in their own words. The model reads each CV against one criterion at a time and returns a level plus the exact line from the CV that supports it. No quote, no level. If the CV says nothing, the answer is unclear: a question for the recruiter, not a penalty.

Every override takes two seconds and is logged with a reason. That log became the most valuable thing in the product: it shows where a rubric is too narrow, and it is the test set every prompt or model change has to pass.

It never rejects anyone by itself. Proposed rejections go into a batch a person actually reads, least certain first. Only when they confirm does an email leave.

Role: solo build, from scope to production

Rubric-based scoring with quoted evidence per criterion

Recruiter overrides logged and reused as an eval set

Human-confirmed batch rejections, built with the EU AI Act in mind