Common questions, honest answers.

Pulls jobs from LinkedIn, Indeed, Glassdoor, ZipRecruiter (via licensed aggregators — no scraping), plus the careers pages of thousands of companies that host on Greenhouse, Lever, Ashby, and Workable (Anthropic, Stripe, OpenAI, every YC company, and more). Then scores every posting against your profile with honest rationale, tailors resume + cover letter for each role, drafts follow-ups, and parses recruiter replies. All inside one product, with hard AI-spend caps you control.

Not by scraping (that's against LinkedIn's terms). We have a Chrome extension that clips jobs from any board you're already logged into — LinkedIn, Indeed, Glassdoor, AngelList, company career pages — and pushes them to NovaHunt for scoring. The clipping happens in your own session, so it's ToS-respecting.

Brutally. APPRAISER undermatches when fit is weak — we'd rather show you a 27 than fake an 87. Every score comes with a plain-English rationale from JUDGE explaining the weighted factors. Then SCOUT (a local scikit-learn re-ranker) personalizes the order based on what you actually save vs dismiss.

Email applications, yes — DISPATCHER drafts and sends with your reply-to, you confirm before send. Full ATS browser automation (Greenhouse/Lever/Ashby/Workable forms) is on the roadmap; today we stage the tailored docs and hand you a deep-link to submit manually.

PERSONA reads the job description and tells you what archetype the strongest candidates fit — their typical experience, skills, certifications, the differentiator that wins the role — then marks which traits you already have based on your resume. Helps you decide which jobs are worth applying to.

Every Claude call passes through TURNSTILE, a deterministic Python validator with zero LLM imports. If your daily cap is $5 and the next call would push you over, TURNSTILE returns 402 'cap reached' and the agent halts. No LLM can reason its way past the gate.

Sourcing and scoring pause until the next billing cycle, or until you top up via a $10 credit pack (no auto-renew on credit packs). We never charge an overage without explicit user action.

One click in the Stripe Customer Portal. Cancellation takes effect at period end — you keep access until then. We refund within 14 days if you've used less than 10% of your included AI budget on the first cycle.

Not yet, but it's on the roadmap. When we ship it, bring-your-own-key bypasses our spend caps entirely — the limits become whatever Anthropic enforces on your account.

When someone signs up through your link and pays for their first month, you both get a $5 credit on your next bill. Credits stack — refer ten friends, pay for nothing. Find your link in Settings → Refer a friend.

No. We never train on your data. Anthropic processes resumes under their zero-retention API setting. We use Claude for inference only — the model never sees your data twice.

Yes. Every document is yours, stored in your account, and exportable as Markdown. We do not retain copies after you delete your account.

Settings → Account & data → Delete my account. One click + email confirmation cascade-deletes every row tied to your account (profile, resumes, jobs, matches, applications, documents, agent runs, notifications, subscriptions, audit events). GDPR Art. 17 + CCPA right-to-delete.

Yes. Settings → Account & data → Download my data. One JSON file with every row tied to your account. GDPR Art. 20 portability.

Only essentials: Auth.js session cookie + CSRF token (required for sign-in) and one optional Sentry cookie for error correlation. No ads, no analytics, no tracking pixels. The opt-out banner lets you disable the Sentry cookie.

A single monolithic LLM is impossible to govern. You can't measure spend per task. You can't cap one capability without breaking another. You can't audit its reasoning. NovaHunt splits the job into 32 small agents organized in three tiers, each independently rate-limited, spend-capped, and swappable.

WARDEN, TURNSTILE, METER, CANARY, ABACUS — written in pure Python with zero LLM imports. They validate every agent action and authorize every Claude call. An LLM literally cannot reason its way past them because they're not LLMs.

It's a local scikit-learn TF-IDF + LogisticRegression model that re-ranks your match list based on every save and dismiss. The more you click, the sharper the ranking gets. Runs on the same CPU as the FastAPI worker — zero additional AI cost per re-rank.

Railway (Postgres + Redis + the FastAPI backend + Next.js frontend). US-region by default. Disk encryption at rest, TLS 1.3 in transit, AES-GCM for our MCP credential vault.