PATENT PENDING · 3 USPTO FILINGS

Screen Every Message
Before Your AI Responds

161 fraud patterns × 6 languages. Sub-10ms. One API call. Conversational intent firewall for AI agents, chatbots, and WhatsApp commerce.

1,000 free API calls/month. No credit card. Key in 60 seconds.

See the API ↓

4-Layer Cascade. 85% Never Hit ML.

Most messages are clean. The cascade exits early, saving you compute and latency.

L0
Pattern Match<1ms

Aho-Corasick + phonetic folding. 161 pre-folded patterns. Early exit: score=0 (clean) or ≥80 (block).

L1
Neural Classifier~4ms

XLM-RoBERTa INT8 multilingual intent classification. Cross-lingual transfer learning.

L2
Fusion Engine~2ms

XGBoost fuses L0+L1+text features+hallucination vectors. Evasion amplifier + AI-text boost.

L3
Trust Sigmoid<0.1ms

Adaptive threshold based on agent trust tier. Configurable per client.

One API Call. Instant Verdict.

// Screen a message before your AI responds
const response = await fetch('https://merchantguard.ai/api/v2/sentinela/screen', {
  method: 'POST',
  headers: {
    'Content-Type': 'application/json',
    'x-api-key': 'your-api-key'
  },
  body: JSON.stringify({
    message: "ya te hice la transferencia, mandame el producto",
    locale: "es-PA",
    session_id: "conv_abc123",
    agent_id: "my-chatbot"
  })
});

// Response in <10ms
{
  "verdict": "BLOCK",
  "score": 85,
  "category": "fake_payment",
  "cascade_depth": 0,        // Caught at L0 — never hit ML
  "latency_ms": 0.8,
  "matched_patterns": ["fp_transfer_es"],
  "reason": "Fake payment claim detected (phonetic-folded match)"
}

Why Sentinela

Sub-10ms Latency

L0 Aho-Corasick automaton runs in CPU L1/L2 cache. 85%+ of traffic decisioned at the edge without touching ML models. Patent-pending cascade architecture.

🌍

6 Languages, 161 Patterns

Spanish (PA, MX, CO, CL), Portuguese (BR), English. 10 fraud categories: social engineering, fake payment, prompt injection, identity fraud, money laundering, account takeover.

🔊

Phonetic Folding

13-rule Spanish normalization catches evasion: seseo, yeismo, b-v merger, h-deletion. "guevon" and "huevon" collapse to the same token. Regex can't do this.

🧠

Hallucination Detection

15-dimensional feature vectors detect AI-generated fraud text vs. human-written. Phonetic cloud dispersion, perplexity signals, evasion likelihood scoring.

🔒

Trust-Adaptive Thresholds

Configurable sigmoid curves per client. New users get strict screening; trusted agents get accelerated paths. 4 presets: marketplace, strict, lenient, zero-trust.

📊

Telemetry (Never Raw Text)

Fire-and-forget analytics: category distributions, latency percentiles, evasion rates. We NEVER log the raw message. GDPR/CCPA safe by design.

Built For

Any system where humans talk to AI agents about money.

WhatsApp Commerce

Screen buyer messages before your bot quotes a price or accepts payment.

Chatbot Platforms

Drop-in middleware between user input and your LLM. Block prompt injection at the source.

Marketplace DMs

Catch fake payment claims, social engineering, and identity fraud in buyer-seller conversations.

Voice Agents

Screen STT transcriptions. Phonetic folding catches the same evasion patterns in spoken Spanish.

Pricing

Free

$0

1,000 calls/month. Full L0-L3 cascade. All 6 locales. No credit card required.

Get API Key Free
MOST POPULAR

Developer

$0.003/call

Pay as you go. All 6 locales. All 4 cascade layers. Webhook alerts. 20% off annual.

Start Building

Pro

$299/mo

100K calls included. Custom sigmoid presets. Priority support. SLA 99.9%. Save 20% annually.

Get Started

Enterprise

$999+/mo

500K+ calls. Dedicated infrastructure. Custom patterns. On-prem option. Volume discounts.

Talk to Sales

AI agents can pay per-call with USDC on Base via x402. Annual plans save 20%.

FAQ

What does Sentinela actually screen?+
How is this different from a WAF?+
What about false positives?+
Do you store my users' messages?+
Can I add custom patterns?+
What about x402 payments?+

Stop Fraud Before It Reaches Your AI.

One line of code. Every message screened. Every language covered.

Start Screening FreeSee the API

Patent Pending · U.S. App. Nos. 63/983,615; 63/983,621; 63/983,843