A CV tool built so it can't quietly lie for you

The problem it exists for

Ask a general-purpose model to tailor a CV to a job description and it does something reasonable-looking and quietly dangerous. Where the posting asks for something your history does not contain, it writes the sentence a strong candidate would have written. A metric. A team size. A tool you have never opened.

The output is fluent, so it is hard to catch on a reread — you are scanning a document about your own career, at the end of a long day, looking for typos rather than for claims. And the place it surfaces is the worst possible one: an interviewer picks the strongest-looking line on the page and asks you to walk through it. An invented number does not survive two follow-up questions, and by then the cost is not a rejection, it is having been the person who wrote it.

That failure is not a bug in a particular model. It is what unconstrained generation is: producing plausible text, where plausibility and truth are different things. So the fix has to be structural rather than a better prompt.

How that shaped the product

Your CV is parsed once into a fact graph — atomic statements about what you did, where, when, with what. Tailoring selects and rephrases from that graph rather than writing from the job description, so a line can change emphasis and wording but cannot acquire a fact the graph does not hold.

Then a separate pass reads the finished draft as though it had never seen your CV, extracts every claim, and labels each one against the graph. Independence is the point: the step that checks is not the step that wrote. The scoring on top of it isdeterministic TypeScript, not a model's opinion — the same CV and job always produce the same numbers, and a truthfulness score below 85 caps the overall score no matter how well the CV matches.

None of that makes fabrication impossible, and it would be a bad sign if this page claimed it did. What it does is make fabrication visiblerather than silent. That is a smaller promise and a keepable one.

Who is building this

Antoniu Gavris — one person, writing the pipeline, the app, the extension and these pages. You can check the work at GitHub and LinkedIn.

Being one person is worth saying plainly rather than hiding behind a "we". It means the product moves quickly and has fewer integrations than any incumbent it is compared against. It also means there is nobody to blame for a claim on this site, which is a useful constraint to work under when the entire pitch is about not overstating things.

What this site will not do

The rules the writing here is held to, because a company arguing that its category is careless with claims should be checkable on its own:

• No invented statistics. No user counts, no case studies, no testimonials, no "candidates see a 40% increase" figures. If a number appears, it comes from the product or from a linked source.
• Competitor claims are quoted and dated. Every row in the comparison pages links to that vendor's own material and carries the date it was read. A vendor's facts failing to be re-checked within ninety days fails the test suite.
• An absence is reported as an absence. When a competitor makes no public statement about accuracy, the page says where it looked and states plainly that this is not evidence they fabricate.
• Every comparison carries a reason to use the other tool.If you are writing a first CV from a blank page, SiviGen is the wrong purchase and its pages say so.

Where it is weak

It tailors a CV you already have and cannot write one from nothing. It cannot audit your source material — if your uploaded CV overstates something, that overstatement is a fact as far as the graph is concerned. The template library is four. TheChrome extension is not on the Web Store yet. And it is young, with less surface area than anything it gets compared to.

The thing it has instead is the mechanism, and a refusal to sell past it.