SHIVA - Guardrails for AI agents — proof, not promises

"If your agent can break Shiva, it better be good at breaking Bitcoin" - AI agents now spend money, send emails, delete data, and ship code — and logs can be edited. Shiva is the accountability layer: every action gets an Allow, Block, or Pause verdict in under 200 ms, sealed into a SHA-256 hash chain, HMAC-signed, and streamed to your own storage. Cryptographic proof for every decision. Bring your own model, your own keys. Walk into any audit with proof — not promises.

Hey Product Hunt 👋 I'm Dheeraj — the person behind Shiva. Former café owner in Rishikesh, India. Went bankrupt. Moved into a broken house in the Himalayas with mice for company. Started learning to code at 43 by asking AI to explain everything like I was 12. I didn't set out to build a governance platform. I was trying to build something small. An AI told me midway through — "you were trying to buy milk and you accidentally built a rocket." That's genuinely what happened. What Shiva actually does: Every action your AI agent takes passes through Shiva before it runs. Keyword filter in microseconds. Local model if needed. Cloud model only if genuinely unsure. Verdict back in under 200ms — sealed into a SHA-256 hash chain with HMAC signatures your agents can never touch or rewrite. Alter any record and the chain breaks instantly. You can verify it yourself in the browser right now. What makes it different from every other guardrail: Most tools log what happened. Shiva prevents it — then cryptographically proves it did. Your data never lives on our side. Every sealed verdict streams straight to your own storage, under your keys. Where it stands: Free right now. Up to 3 agents, 1,000 evals each. EU AI Act high-risk obligations kick in August 2, 2026 — five weeks away. Ungoverned agents are a liability now, not later. I read every message personally. If something breaks, I want to know. If you think the threat model has holes, tell me — I'll think through it with you. Happy to get into the technical details, the cryptographic design, or where it currently falls short. — dkbk

SHIVA - Guardrails for AI agents — proof, not promises

Replies