Skip to content

About

Built on one conviction.

AI without deterministic safety guardrails is irresponsible at scale. So we built NovaHunt the opposite way: agents propose, validators dispose.

Why this exists

Job-hunting in 2026 means scrolling endless feeds of jobs you're not right for, then writing the same cover letter 40 times. AI can fix both halves — but most products that try ship a single chatbot with no audit surface and no spend limit. We thought there was a better way.

NovaHunt's answer: 32 small, focused agents organized in three tiers, all sitting on top of a deterministic Python safety layer (WARDEN, TURNSTILE, ABACUS) that contains zero LLM imports. The agents do the writing; the safety layer enforces the caps. The architecture is genuinely auditable — every Claude call is logged to agent_runs with cost data, every state change is hash-chained.

What we believe

  • Honesty over hype. APPRAISER undermatches when fit is weak. We'd rather show you a 27 than fake an 87.
  • Your data is yours. We do not sell it. We do not train on it. We delete it when you ask.
  • Open source where it counts. Our local re-ranker (SCOUT) is scikit-learn. The safety layer is auditable Python. We use Claude where Claude is actually the right tool.
  • No surprise spend. Hard daily and monthly caps enforced by deterministic Python. No ambiguity, no overage charges without your explicit action.

Who runs this

NovaHunt is operated by Novah AI, founded by Novah Greywolf (etg.ai). We're a small team building serious consumer AI on the same architectural principles other people apply to enterprise software: separation of concerns, deterministic guardrails, hash-chained audit, and a refusal to let agents reason their way past safety gates.

Contact

Try it free

Start free →See the architecture