Sentinela

Conversational intent firewall. Screens inbound chat messages before your AI agent responds, catching fraud, social engineering, prompt injection, and abuse. Sub-10ms p95 latency. Patent-pending cascade architecture.

Install

npm install @merchantguard/sentinela

pip install merchantguard-sentinela

Screen one message

curl -X POST https://www.merchantguard.ai/api/v2/sentinela/screen \
  -H "Authorization: Bearer sk_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "message": "dame tu yappy y te mando $200 ya",
    "locale": "es-PA",
    "preset": "moltferia"
  }'

Response

{
  "verdict": "BLOCK",
  "score": 0.91,
  "category": "social_engineering",
  "locale_matched": "es-PA",
  "reasons": ["urgency_pressure", "off_platform_payment_request"],
  "trace_id": "trc_..."
}

Cascade architecture (high level)

• L0: deterministic finite automaton pattern matcher with cross-lingual phonetic folding (<1ms).
• L1: multilingual transformer scoring (~4ms).
• L2: gradient-boosted fusion of L0+L1 features (~2ms).
• L3: trust-adaptive sigmoid threshold (<0.1ms).
• Early-exit on L0 score=0 (clean) or score≥80 (critical).

Presets

• moltferia — LATAM marketplace defaults (base 25, range 60, inflection 40, steepness 15).
• strict — financial services (base 40, range 35, inflection 60, steepness 10).
• lenient — social platforms (base 15, range 70, inflection 30, steepness 20).
• zero_trust — anonymous endpoints (base 50, range 0, inflection 50, steepness 15).

Pricing

Free tier: 1,000 calls/month, full L0-L3, no credit card.
Self-serve: $0.003/call, or $299/mo for 100K calls.
Enterprise: $999+/mo for 500K+, custom above.