About our AI

What the AI does

• Match-gate scoring. Reads the job posting and your evidence, returns a score (0-100) with a band label (Below gate / Moderate / Strong / Exceptional) and a reasoning paragraph. Below 50 it tells you the gap is real and stops — that's a correct outcome, not a failure.
• CV draft. Tailors your existing CV to the posting. Every claim about your experience cites the upload it came from — if no source exists, the AI asks you instead of inventing.
• Cover letter draft. Written in your voice (drawn from your sample writing), not AI-default. The drafted letter should read like you wrote it yourself after a good night's sleep.
• Application email draft. The short message that accompanies CV + letter. Sent from your address, tracked in your applications view.
• Coaching exercises (values, motivators, qualities, STAR-R). Pre-populated cards you swipe / sort / pick from. The AI helps you crystallise what you already know — it doesn't tell you who you are.

What the AI does NOT do

• Decide for you. Every draft is a starting point — you read, edit, approve, send.
• Learn from your data to improve anyone else's experience. Your data does not train any model — ours or anyone else's.
• Talk to recruiters or employers on your behalf without your explicit click-to-send.
• Invent claims about your experience. Reframing a requirement to make it fit is treated as a gap to surface, not a writing opportunity.
• Replace human review for high-stakes decisions. Hiring is high-stakes. We never claim the AI's match-gate score is an authorisation to apply (or not apply).

How we verify the AI is honest

Every AI output goes through a 7-layer safety chain BEFORE it reaches you:

1. Persona-shift detection. Catches the model lapsing into 'as an AI…' / 'I cannot…' / 'in summary' boilerplate.
2. Citation enforcement. Every factual claim about you MUST cite a source in your own evidence. Uncited claims are flagged.
3. Output scrubbing. Banned-phrase patterns (per language) get redacted.
4. Canary detection. Hidden markers in the input let us detect if an attacker tried to inject instructions via a job-posting URL or an inbound email.
5. PII scan. Catches personal data the model shouldn't be emitting.
6. Input screen. Filters jailbreak attempts BEFORE we ask the model anything.
7. Indirect-injection scan on inbound mail. The reply-tracker filters injected instructions from recruiter emails.

When any layer flags, we log the incident to our regulator-facing audit log (per EU AI Act Article 26(5)) and surface a warning on the affected page.

How you stay in charge

• Every AI-generated surface carries a clear AI badge (the small pill in the top-right of each draft card) so you know which content is AI-drafted and which is your own.
• Every numeric AI score carries a band label and tooltip explaining what the number means + reminding you it's a calibration estimate, never a decision.
• Every claim in an AI draft carries a [src: …] citation — click it to inspect the underlying evidence in your uploads.
• If the AI couldn't form a confident answer, it surfaces a "Not sure" panel instead of pretending. Honest "I don't know" beats a confidently-wrong draft.
• Every time an Omoikane operator reads your data, it is recorded in your personal access log (read it anytime).
• You can export everything anytime, and delete everything anytime, from settings. Deletion cascades within 7 days.
• The short notes you'll see across the app — that the AI's output is a data point, not a decision — are there to counter automation bias. That page explains why, and how to read AI output well.

Which AI provider runs the model

On the managed tiers, your requests go to an operator-selected provider — a single platform setting, changeable without a release, which is why this page names a mechanism rather than a brand. Which provider is active, and where it processes data, is listed on the Sub-processors page. You can also bring your own LLM key (Bring-Your-Own-Model / BYOM): Anthropic, OpenAI, Mistral, DeepSeek, NVIDIA NIM, Ollama Cloud, or any OpenAI-compatible endpoint. You configure this at /profile/ai-providers; we never silently fall back from your key to ours.

Every provider has a data-residency disclosure banner shown at the moment you add the key — DeepSeek = PRC, Anthropic/OpenAI/NVIDIA NIM = US, Mistral = EU, Ollama Cloud = provider-stated, generic OpenAI-compatible = user-verified.

The formal policy

This page is plain-language. The binding documents are:

• Privacy policy — what we collect, why, how long we keep it, how to delete.
• Terms of service — what we promise, what we ask in return.
• Cookie policy — three first-party cookies + the cookieless Umami analytics caveat.
• Sub-processor register — the full list of companies whose code touches your data.