# Building an AI CV tool that refuses to invent experience > Source: > Published: 2026-08-11 19:47:48+00:00 Disclosure: I built [CV Aligner](https://cv-aligner.com). This is a technical product note, not an independent review. A common failure mode in AI-assisted CV writing is not obviously absurd output. It is a sentence that sounds credible, reads well, and quietly says more than the source material supports. I wanted to build around that failure rather than add another general-purpose rewrite button. CV Aligner compares a CV with a real job description requirement by requirement. For each requirement, it shows the source lines supporting the assessment and suggests reframes that remain tied to those lines. Employers, dates and metrics are locked. If the CV contains no evidence for a requirement, the product should expose the gap rather than manufacture a plausible substitute. That creates a useful product boundary: The front end uses React, Vite and TypeScript. Anthropic models are called server-side so the application can enforce structured inputs and keep model access out of the browser. The model is only one part of the system. The more important work is deciding what data it may change, what must remain immutable, and what evidence must accompany each suggestion. The evidence report is free and requires no account or payment. Optional application packs are one-off purchases rather than subscriptions. I do not yet have first-party evidence that CV Aligner improves interview or callback rates. It would be easy to imply that a better-aligned document automatically produces a better hiring outcome, but the evidence is not there yet. The narrower claim is testable: the tool shows how each suggestion relates to the source CV and makes unsupported gaps visible. You can try the free evidence report at [cv-aligner.com](https://cv-aligner.com). For developers building constrained AI products, where have you found that limiting the model improves trust, and where does it simply create friction?