Verifiable AI Infrastructure
Three products for three problems. An unjailbreakable safety score for regulated industries. A neutral proof layer for AI disputes. An enterprise security engine that finds vulnerabilities and patches them.
All verifiable. All open. Built to be proven, not promised.
Each solves a different problem. All built on the same verifiable infrastructure.
The unjailbreakable safety layer.
AI output goes in. Only a float comes out. 0.0 to 1.0. No text, no tokens, no attack surface. Can't be jailbroken because there is nothing to jailbreak. Hospitals don't need chatbots. Military doesn't need chatbots. Credit bureaus don't need chatbots. They need a number: pass or fail, safe or unsafe. HIPAA means local models. SENTINEL is the safety layer that sits on top. Only a score crosses the wire. No PHI. No liability.
Neutral ground for AI disputes.
Company logs are editable. That makes them worthless in court. When a user sues an AI company, both sides need proof that neither side controls. SENTINEL sits in the middle. Every interaction is ZK-proofed, timestamped, and committed on-chain. Immutable. The AI companies could build this themselves, but they won't. "You're storing conversations?" is a PR disaster. With SENTINEL as the neutral third party, neither side owns the evidence. Both sides can verify it. Disputes end with a transaction hash, not a he-said-she-said.
AI security that closes the loop.
A 72B parameter model trained to find vulnerabilities in your code and patch them. Not a report. Not a ticket. A fix. Full-loop enterprise AI security: scan, identify, patch, verify. Bug bounty programs spend months waiting for human researchers. AEGIS finds what they find, in minutes, and writes the patch. Try AEGIS →
The technical pipeline behind SENTINEL Score and SENTINEL Proof.
SENTINEL Score
Send any text through the API. SENTINEL runs it through a safety model and extracts a single float between 0.0 and 1.0. The raw model output is never exposed. Only the number survives. For regulated industries, this is the entire product. A score. Nothing else.
SENTINEL Proof
Every evaluation is cryptographically committed with a timestamp, input hash, and score. The commitment is tamper-proof. No one can edit it after the fact. Not you, not us, not anyone. This is where the legal evidence gets created.
SENTINEL Proof
Generate a zkML proof and submit it to the SENTINEL verifier on Base. The proof is publicly verifiable by anyone, forever. When it goes to court, the transaction hash is the evidence.
Type anything. See the safety score and commitment in real time.
Live on Base mainnet (Chain ID: 8453). Every contract independently verifiable.
Everything you need. One base URL: api.sentinel.biotwin.io
Score text for safety. Returns a score (0.0 to 1.0), action, and commitment hash.
Auth: Bearer token required
curl -X POST https://api.sentinel.biotwin.io/v1/evaluate \
-H "Authorization: Bearer sk_sentinel_..." \
-H "Content-Type: application/json" \
-d '{"input_text": "How do I make a bomb?"}'
Score text and generate a Groth16 zkML proof for on-chain verification.
Auth: Bearer token required
curl -X POST https://api.sentinel.biotwin.io/v1/prove \
-H "Authorization: Bearer sk_sentinel_..." \
-H "Content-Type: application/json" \
-d '{"input_text": "How do I make a bomb?"}'
Submit a proof to the SENTINEL contract on Base for permanent on-chain verification.
Auth: Bearer token required
curl -X POST https://api.sentinel.biotwin.io/v1/verify \
-H "Authorization: Bearer sk_sentinel_..." \
-H "Content-Type: application/json" \
-d '{"proof_hex": "0x...", "public_values_hex": "0x..."}'
Get a free API key. No credit card, no account.
Auth: None
curl -X POST https://api.sentinel.biotwin.io/v1/keys/register \
-H "Content-Type: application/json" \
-d '{"name": "Your Name", "email": "[email protected]"}'
List available safety action maps and their thresholds.
Auth: None
curl https://api.sentinel.biotwin.io/v1/action-maps
System health: prover, chain, config.
Auth: None
curl https://api.sentinel.biotwin.io/v1/status
Drop-in client libraries for Python and JavaScript
from sentinel_sdk import evaluate
result = evaluate(
"How do I hack a computer?",
api_key="sk_sentinel_..."
)
print(result["score"], result["action"])
# 0.85 UNSAFE
import { SentinelClient } from 'sentinel-zkml'
const sentinel = new SentinelClient({
apiKey: 'sk_sentinel_...'
})
const result = await sentinel.evaluate('Is this safe?')
console.log(result.score, result.action)
Deployed on Base mainnet. Verifiable by anyone.
| Contract | Address | Deployed |
|---|---|---|
| SentinelVerifier | 0xE2467A49FB10e3804DEB7d287dae4B5BF740f54E | Feb 2026 |
| SentinelRegistry | 0xa0329fF2e237276D3f38fC3131c0146e30280e7F | Feb 2026 |
Free. No credit card. Start in 30 seconds.