I built Vekt, and the first thing we published was a list of what it can't do. Here's why.

Vekt is an open-source tool that screens CVs against a job description and gives recruiters a ranked list of who to read first. It runs on your own servers, and with a local model, the CV never leaves them.

In the EU, screening CVs with AI is classed as high-risk under the AI Act, and GDPR already gives candidates rights over automated decisions. It's easy to answer that with a "compliant" badge. We wanted something you could check instead.

So before the feature list, the repo has a compliance folder that says plainly:

• No accuracy has been measured yet.

• Redaction removes names and contact details, not every clue about who someone is.

• Scores from different models can't be compared.

• Running it yourself doesn't make you compliant.

What it does do:

• Removes direct identifiers before any text reaches the model

• Scores each application with short reasons, pros and cons

• Keeps a person in charge: auto-reject is off by default, and status emails wait 48 hours so a recruiter can change the call

• Shows candidates that AI is involved and how to ask for human review

• Works with OpenAI, OpenRouter or a local model via Ollama

• MIT licence, free to self-host

The first item on that list is the biggest gap. So today's ask isn't a discount: we're looking for 2–3 organisations to run a proper accuracy and bias evaluation with Vekt on their own servers. We'll help set it up, and with your agreement the method and results get published.

A question for you: if you hire with AI today, what would you need to see before you trusted a score?