AI-assisted fixes
AI explains and suggests. The score is deterministic and AI never touches it.
“Questions about the report” opens after the analysis.
- Local model (Layer 1): Embedding-based semantic matching that runs in the browser.
- Your own Ollama (Layer 2): A model on your machine for bullet rewriting, explanations, cover letters and interview questions.
- BYOK (Layer 3): Stronger models with your own API key. The key stays in the browser and is never sent to our server.
- Schema validation: Every model output is checked against a JSON schema. Output that fails is never shown.
Every AI output passes a grounding check that looks for invented numbers, technologies or organisations that are not in the CV.
Does AI feedback change my CV score?
No. The AI feedback in ATS Free For All explains findings and suggests edits, but the CV score out of 100 is calculated only by fixed rules, and no model output is ever fed into it.
The scoring engine is deterministic: the same resume text always gets the same score. The AI code runs in a separate part of the app and has no path into scoring. If you accept a suggested rewrite, the edited text is measured again by the same rules, so any change in the score comes from the rules reading the new sentence, not from a model judging it. For bullet fixes, a rule-based draft is tried first, and a model is asked only when no rule applies. The model selector starts on No model, and the full report, fix list and keyword view work without any AI at all.
Which AI models can I use, and where do they run?
ATS Free For All offers four optional text models plus No model, and three of them run on your own device: the browser's built-in model, WebLLM inside the browser, and Ollama on your computer.
A text model is used to explain a finding, rewrite weak bullets, draft a cover letter, generate interview questions from a job ad and answer questions about your report. Separately, an optional semantic hint model (multilingual-e5-small, about 118 MB) runs in the browser and points out near-misses the literal matcher cannot see, such as a resume saying Postgres where the ad says PostgreSQL. Nothing downloads until you agree, the size is shown first, and the download can be cancelled.
| Option | Where it runs | What to know |
|---|---|---|
| Browser built-in | Chrome or Edge, on your device | Gemini Nano; English and German only |
| WebLLM | In the browser over WebGPU | One-time download of about 1600 MB; needs WebGPU |
| Ollama | Your own computer (localhost) | Contacted only when you select it; needs OLLAMA_ORIGINS set for this site |
| Bring your own key | The provider you configure | Key stays in your browser; you approve every request |
| No model | Nothing runs | Deterministic report only; the default |
How is an AI rewrite checked? An example
Every AI rewrite in ATS Free For All must pass a schema check and a grounding check, and a sentence that adds a number, technology or organisation missing from your original bullet is discarded instead of shown.
Example: your bullet reads Responsible for migrating the billing service to Kubernetes. A rewrite such as Migrated the billing service to Kubernetes, [quantify: downtime or cost impact] is accepted, because the placeholder marks what only you can fill in. A rewrite such as Migrated the billing service to Kubernetes, cutting costs by 30% is dropped, because 30 does not appear in the source. If the model returns the wrong structure twice, you see an error rather than a guess. An explanation request sends the model only the finding and its evidence lines, and a question about the report sends scores, findings and term lists, not the full CV text.
What do the AI features not do?
The AI features in ATS Free For All do not change scores, do not run until you choose a model, and do not send CV text to an outside provider unless you set up your own key and approve the request.
With Bring your own key, the key is kept in your browser, and before each request a confirmation names the provider's host; declining stops the request. The browser built-in model handles English and German only, so Turkish requests need WebLLM or Ollama. WebLLM needs WebGPU to run at a usable speed. A grounded rewrite can still be weak or awkward, so read every suggestion before you apply it. The AI also cannot tell you whether an employer will invite you to interview: the score measures how readable the document is, not your chances.
“Questions about the report” opens after the analysis.
Last updated