Releases: Siddharthpatni/Sentinel
Releases Β· Siddharthpatni/Sentinel
Sentinel v0.1.0 β Phase 3: Observability, Verifications, Routing, Evals, Audit & Alerts
π‘οΈ Sentinel v0.1.0 β first public release
Sentinel is an open-source, self-hostable LLM observability + agent reliability platform. Drop-in proxy for OpenAI / Anthropic that gives you traces, evals, routing, EU AI Act audit logs, and alerts β with zero code changes beyond two import lines.
β¨ What's included in this release
- Observability β every LLM call traced with cost, latency, tokens, and full request/response bodies in a live dashboard. 24h cost sparkline included.
-
- Verifications β declarative rules re-check primary calls with a judge model. Sampled, async, never blocks the caller.
-
- Routing & fallback β per-request model overrides with ordered candidate fallback (up to 3 attempts). Streaming bypasses preserved.
-
- Evals β YAML-defined regression suites with 7 assertion types (contains / equals / regex / max-latency / max-cost / json-schema / llm-judge), run-history UI, pass-rate trend endpoint, and a CI entrypoint for GitHub Actions.
-
- EU AI Act audit log β risk-tier classifiers tag inbound calls; every tagged call lands in a SHA-256-chained ledger an auditor can verify offline. NDJSON export, server-side
/verify.
- EU AI Act audit log β risk-tier classifiers tag inbound calls; every tagged call lands in a SHA-256-chained ledger an auditor can verify offline. NDJSON export, server-side
-
- Alerts β threshold checks on cost-per-hour, error-rate, and p95 latency over rolling windows.
-
- BYOK β per-project provider keys (OpenAI / Anthropic / OpenRouter / Gemini), encrypted at rest with Fernet, live-validated against provider model-list endpoints.
-
- Python SDK β
pip install sentinel-sdk. Drop-inOpenAI/Anthropicclients plus a programmatic control plane for rules, routing, and evals.
- Python SDK β
π Quick start
git clone https://github.com/Siddharthpatni/Sentinel.git
cd Sentinel
cp .env.example .env # fill in OPENAI_API_KEY / ANTHROPIC_API_KEY
docker compose up -d
# Open http://localhost:3000 β traces appear in real timeπ Architecture
FastAPI gateway Β· Next.js dashboard Β· PostgreSQL Β· Redis + Celery worker. MIT-licensed, self-hostable, production-shaped.
π Docs
docs/verifications.mdΒ·docs/routing.mdΒ·docs/evals.mdΒ·docs/audit.mdΒ·docs/alerts.md-
- Conceptual learning notes in
docs/learn/
- Conceptual learning notes in
π¬ Feedback
This is the first public cut β feedback, issues, and PRs are very welcome. If you're shipping LLM features in production and have a war story about agent reliability, I'd love to hear it.
License: MIT