rbrtbmnglg/guardrailsanalysis

Guardrails360 — AI System Instruction Auditor

Paste any AI agent system prompt and get a structured audit report of which security, privacy, and governance guardrails are present or missing — with severity ratings and remediation guidance.

What it does

Guardrails360 runs your system instruction through 7 specialized auditor agents built on CrewAI. Each agent inspects the prompt against a specific compliance domain and returns structured findings. A governance agent then synthesizes everything into a single JSON report with per-guardrail status, severity, and remediation steps — streamed to the UI in real time.

Audit categories

Each guardrail in the report carries a status (PRESENT / MISSING), severity (Critical / High / Medium / Low), a description of the mechanism, and concrete remediation guidance.

Optional features

• —Gatekeeper — pre-flight LLM classifier that rejects non-prompt input before the full crew runs
• —Cost / latency profiling — tier breakdown estimating per-1k-request compute cost
• —Green-AI scoring — estimated energy per 1k requests with an efficiency score
• —Advanced reasoning — switch the analysis engine between DeepSeek-V3, Llama-3.3-70B, and Qwen2.5-72B

Usage

1. 1.Open the live Space
2. 2.Paste your system instruction (10–20 000 characters)
3. 3.Enter a HuggingFace API token with Inference access (hf_…)
4. 4.Toggle any optional features
5. 5.Click Analyze — results stream in as each agent completes

Your token is used client-side to call the HF Inference Router; it is never stored on the server.

Architecture

• —Backend: FastAPI serving a single POST /analyze endpoint (NDJSON stream)
• —Agents: CrewAI sequential process — 4 audit agents run, then optional cost/green-AI agents, then governance synthesizes
• —LLM routing: HuggingFace Inference Router (OpenAI-compatible), user-supplied token
• —RAG: OWASP LLM Top-10 PDF indexed with sentence-transformers + ChromaDB
• —Frontend: Vanilla JS (6 modules), Tailwind CSS, real-time progress bar
• —Deploy: Docker on HF Spaces (port 7860)

Local development

# 1. Clone and install
git clone https://github.com/trbbmnglg/guardrailsanalysis
cd guardrailsanalysis
python -m venv .venv && source .venv/Scripts/activate  # Windows: .venv\Scripts\Activate.ps1
pip install -r requirements.txt

# 2. Run
uvicorn main:app --reload --port 7860
# Open http://localhost:7860 — the UI will ask for your HF token

Deployment

Pushing to main on GitHub triggers .github/workflows/sync-to-hf.yml, which force-pushes to the HF Space and triggers an automatic Docker rebuild. GitHub is the single source of truth — never push to HF directly.

License

MIT