Sentryn - AI Agent Guardrail Platform

🟢 Live Demo: https://sentryn-engine.ambitioushill-87ee6da5.koreacentral.azurecontainerapps.io 📖 API Docs: https://sentryn-engine.ambitioushill-87ee6da5.koreacentral.azurecontainerapps.io/docs

The Problem Sentryn Solves

As AI agents become more autonomous, a single reasoning error can cause an agent to enter a recursive loop:

1. Agent detects a database issue
2. Agent attempts a fix → fix fails → agent retries with the same reasoning
3. Agent repeats hundreds of times

Result: runaway API costs, infrastructure damage, data corruption. Sentryn stops this before it happens.

Architecture

AI Agent → Python FastAPI Engine (8000) → Redis + Qdrant

• FastAPI Engine — intercepts every agent action, converts reasoning into 768-dim semantic vectors using nomic-embed-text-v1.5, runs kinematic loop detection
• Redis — per-agent execution history with 7-day TTL
• Qdrant — vector embeddings for permanent audit trail
• Azure Container Apps — independently scaled containers in Korea Central

How Loop Detection Works

Every agent action is converted into a semantic vector using nomic-embed-text-v1.5 (768-dim). Sentryn tracks these vectors using kinematic equations:

• Velocity = semantic distance between consecutive actions / time delta
• Acceleration = change in velocity / time delta
• Threat Mass = weighted destructive action scores with exponential decay

When velocity drops near zero (orbital lock), the circuit breaker trips → HTTP 423.

Security

• API Key Auth — X-API-Key header required on all /evaluate requests (401 without it)
• Pydantic v2 Validation — blocks malformed inputs before reaching the model
• Credential Scrubbing — redacts passwords, API keys, tokens before vector storage
• Injection Blocking — 8/8 known injection patterns blocked (jailbreak, DROP TABLE, UNION SELECT, etc.)

Tech Stack

Layer	Technology
API Framework	Python · FastAPI · Uvicorn
Embeddings	nomic-embed-text-v1.5 (768-dim)
State Storage	Redis 7.2
Vector Audit Trail	Qdrant v1.9
Input Validation	Pydantic v2
Cloud	Azure Container Apps (Korea Central)
CI/CD	GitHub Actions → Azure Container Registry
Containers	Docker

Live API

# Health check
curl https://sentryn-engine.ambitioushill-87ee6da5.koreacentral.azurecontainerapps.io/health

# Evaluate agent action
curl -X POST https://sentryn-engine.ambitioushill-87ee6da5.koreacentral.azurecontainerapps.io/evaluate \
  -H "Content-Type: application/json" \
  -H "X-API-Key: sentryn-prod-key-2026" \
  -d '{"action_verb": "DELETE", "target_resource": "prod-database", "reasoning_context": "fixing error", "agent_id": "agent_001"}'

Test Results (26/26 passing)

pytest tests/test_engine.py -v

Category	Tests	Result
API Key Authentication	3	✅
Health & Root endpoints	3	✅
Pydantic validation	4	✅
Injection blocking	8	✅
Credential scrubbing	4	✅
Circuit breaker logic	4	✅
Total	26	✅ 26/26

Benchmark Results (Live Azure)

Metric	Result
Loop detection accuracy	100%
Circuit breaker trips on	2nd repeated action
Injection patterns blocked	8/8
Auth enforcement	401 without key
All backends	Redis ✅ Qdrant ✅ nomic ✅

What I Learned

• Silent exception swallowing hides critical bugs in production — always log warnings
• Azure Container Apps internal DNS resolves short service names but some libraries need prefer_grpc=False
• Mocking heavy ML dependencies (torch, sentence-transformers) enables fast pytest runs without loading 2GB models

Future Improvements

• Next.js dashboard for real-time agent monitoring
• Go Fiber gateway for credential scrubbing at ingress
• Threshold tuning based on empirical loop detection data
• Multi-tenant API key management
• Alert webhooks when circuit breaker trips