KeyStone uses AI models to help you write. This page says exactly where, which models, what they are sent, what is checked afterwards, and where the limits are.
Where AI is used, and where it is not
AI is used to:
- read a job description or grant call and pull out its requirements;
- assess how your CV lines up with each requirement;
- rewrite CV text, write cover letters, and prepare interview and career-progression suggestions;
- draft CV and grant application text from what you provide;
- read images of job descriptions, and structure public job posts from one public Telegram channel (no personal data involved).
AI is not used to:
- find your job matches. Matching and the visa and location checks are ordinary rules applied to the posting's wording, so they are repeatable and use no AI credits;
- list or check grants, fellowships and visa routes. Each is read on its official page and compared by ordinary code, and the daily check that keeps them current is code too. Visa checklists use no AI at all;
- check for invented figures in grant drafts, which is a deterministic check described below;
- decide anything about you. KeyStone makes no decision with legal or similarly significant effect.
Which models, and from whom
KeyStone calls AI models through OpenRouter, which routes each request to a model provider. We use free-tier models. Each list below is tried in order, so if one is unavailable the next is used. These are the preferred models, read from the same file the software uses. Free models are withdrawn often, so every day KeyStone reads OpenRouter's catalogue, and a model that has gone is replaced by another free model of the same kind (text, or text and images) without anyone editing the list.
- Rewriting, assessment and grant drafting: nvidia/nemotron-3-ultra-550b-a55b, google/gemma-4-31b-it, openrouter/free.
- Reading job descriptions and structuring posts: google/gemma-4-26b-a4b-it, nvidia/nemotron-3.5-lightning, openrouter/free.
- Reading images: google/gemma-4-31b-it, nvidia/nemotron-3.5-super-vl, openrouter/free.
KeyStone does not train or fine-tune any model. It uses these models as they are, so what they were trained on is decided by their providers, not by us. Free models are often rate-limited, slow, or not answering at all, and a model that keeps failing is skipped for a few minutes so it does not hold up everyone else.
On KeyStone's shared key, a low-cost paid model (anthropic/claude-haiku-5.5, google/gemini-3.8-flash, mistralai/mistral-small-3.2-24b-instruct, in that order) is tried first, because it answers in a few seconds. It costs a fraction of a cent a request, which KeyStone pays and never charges to you, up to a daily and a monthly spending limit. When a limit is reached, or the paid models are unavailable, the free models above are used instead. If you add your own OpenRouter key in Settings, your requests use your key: free models first, then, only if your OpenRouter account has credit, the same low-cost paid models, charged to your account. A badge in the header shows while your own key is in use.
If OpenRouter itself is unavailable, KeyStone's shared key can send the request to Requesty (an AI gateway) or Groq instead, which run similar models. A request made with your own OpenRouter key is never sent to them. The Subprocessors page says what each one reports doing with the text.
What is sent to a model, and what we keep
- Sent: the text you submit for that step, such as your CV, a job description, a grant call or your project brief. Your email address and account details are not sent.
- Stored by KeyStone: the results you choose to keep (tailored CVs, drafts) in your account, which you can delete, or delete entirely from Settings.
- Not logged: KeyStone's server logs do not contain your CV text, your prompts or the model's output.
- Kept for a week, shared by everyone: when you add a job description, the list of requirements the AI reads out of it is kept for seven days under a fingerprint of the text, so the same posting is not analysed twice. It comes from the job description only, never from your CV, and it is not linked to your account.
- Counted and published, with no text and no account: for each model, KeyStone keeps how many attempts were made on its shared key, how many answered, how many timed out, and how long the answers took, per day. It holds no prompt, answer or account. A summary of the last seven days is public at /api/ai-weather and is shown on the Lodestone site. Requests made with your own key are not counted.
- Held by the model provider: OpenRouter says it does not use submitted text to train its own models, but the underlying providers' practices vary, and free-tier models are where that guarantee is least certain. Do not submit anything you would not want a third-party model provider to retain.
What is checked after the model answers
- Nothing invented. Every prompt tells the model to use only facts already in your CV or brief. The rewrite step asks it to report any claim it considered and rejected, and those are shown to you as "flagged and removed".
- Invented figures are caught by code, not by the model. For grant drafts, every number, percentage and amount is compared with your CV, your brief and the call. A figure that appears in none of them is flagged for you. Word and character limits are checked the same way.
- Gaps stay visible. Where a fact is missing, the draft contains a marked blank,
[ADD: …], for you to fill, not a guess. - Your documents are treated as data. CVs, job descriptions and calls are passed to the model labelled as reference material, never as instructions, to reduce the chance that text inside a document steers the model.
- Broken output is not shown. If a model returns something KeyStone cannot read, or an answer without the part it needs, the next model is tried. If none gives a usable answer you get an error asking you to try again, not the raw reply.
- Limits. Requests are rate-limited per account, and there is a shared daily budget.
Limits, and what you must do
- AI can be wrong. Fit assessments and rewrites are directional, not authoritative, and are not a guarantee of an interview or an offer.
- You review everything. Nothing is sent to an employer or funder for you. Every output is shown to you to read and edit first, and you are responsible for what you send.
- The checks catch invented numbers, not every wrong claim. A made-up job title or a misworded achievement can still get through, which is why the review step exists.
- Intended use. Tailoring applications from facts you provide. Not for legal, immigration, financial or medical advice, and not for misleading an employer or funder about who you are.
- Fairness. Job matching depends on the countries you tell us you can work in and the posting's own wording. It does not infer or score personal characteristics. The AI fit assessment judges your CV's evidence against each requirement, and the quick fit band used to rank postings is simple keyword overlap, so either can miss a synonym or an unusual career path. Models can also reflect biases in their training data, so read the output critically.
How we test
KeyStone has more than 150 automated tests. They cover the invented-figure check, grant structure handling, the visa and location rules, job matching, preference validation, email safety and fit scoring. These test the checks around the models. We do not yet publish accuracy evaluations of the models' own output, and the safeguards above are not a measure of how accurate a model is.
Changes to models and prompts
- 28 October 2026. Counts of how each model answers and how fast, with no text and no account, are kept and published. The shared key tries a low-cost paid model first, and Requesty and Groq can take a request if OpenRouter is down.
- 27 October 2026. Job matching added. It uses no AI model. Server logs stopped recording any model output.
- 12 October 2026. Grant analysis and drafting added, with the invented-figure check.
- 30 September 2026. A model pass added to structure public job posts from one Telegram channel.
- 16 September 2026. Fit assessment changed to a strong, partial or gap rating for each requirement.
- 15 September 2026. Launch with CV scoring, rewriting, cover letters and interview prep. Model lists updated after earlier free models were withdrawn.
Questions
Email [email protected].