v1.0.0 — SentinelLM Initial Release

SentinelLM v1.0.0

First stable release of SentinelLM — real-time safety and quality middleware for LLM applications.

What's included

7 evaluators across 2 layers

• Input: PII detection (redact or block), prompt injection detection, topic guardrail
• Output: toxicity scoring, relevance scoring, hallucination detection, faithfulness scoring

Core features

• Drop-in OpenAI-compatible proxy (/v1/chat/completions)
• 4 LLM backends: Ollama, OpenAI, Anthropic, Gemini
• Concurrent input chain with first-block short-circuit
• Shadow mode for zero-risk threshold tuning
• Redis caching of input evaluator scores
• Fail-open guarantee on all evaluators
• Human review queue for flagged responses
• Real-time WebSocket feed to dashboard
• Eval pipeline CLI with regression detection

Infrastructure

• FastAPI + asyncpg + Redis
• PostgreSQL storage with full request/score history
• Next.js dashboard with live charts
• Prometheus metrics
• Docker Compose (dev + prod)
• 215 unit tests, 61% coverage, mypy clean