Guardrails Enterprise — AI Guardrails & ML Infrastructure

Enterprise AI security and ML infrastructure platform. Wraps any enterprise LLM (healthcare, banking, legal, HR) with input and output guardrail layers, backed by a regulatory RAG corpus.

📁 Repository Structure

guardrails-enterprise/
├── docs/                    # All documentation
│   ├── README.md           # PII NER pipeline documentation
│   ├── QUICK_START.md      # Quick start guide
│   └── SETUP_GUIDE.md      # Detailed setup instructions
├── config/                  # Configuration files
├── dags/                    # Airflow DAGs
├── plugins/                 # Airflow plugins
├── scripts/                 # Utility scripts
│   ├── setup_airflow.sh
│   ├── start_airflow.sh
│   └── stop_airflow.sh
├── docker/                  # Docker configuration
│   ├── Dockerfile
│   └── docker-compose.yml
├── gateway/                 # Phase 1: Input guardrail (3 classifiers)
├── rag/                     # Phase 2: RAG pipeline (Qdrant + Qwen3-4B embedder)
├── multi_agent/             # Phase 3: MAD output guardrail (active)
├── multi_agent_debate/       # MAD pipeline with SQLite storage for GRPO
├── confidence/              # Phase 4: Confidence Scoring Engine
├── rlhf/                    # Phase 5: GRPO feedback loop
├── finetuning/              # Fine-tuning scripts for all 4 models
├── synthetic_data/          # Synthetic evaluation dataset generation
├── MAD_SETUP_GUIDE.md       # Standalone MAD setup guide
└── requirements.txt         # Root-level shared dependencies

🚀 Quick Start

For detailed setup instructions, see:

• Quick Start Guide - Get started in 15 minutes
• Setup Guide - Comprehensive setup instructions
• Full Documentation - Complete project documentation
• MAD Setup Guide - Multi-Agent Debate service setup
• GRPO LoRA Deployment Handoff - Agent A/B adapter location, vLLM dynamic LoRA serving, AWS routing, runtime prompts, JSON schema, and eval summary

✅ Active Pipelines

PII NER Pipeline

• Location: dags/pii_ner_pipeline.py
• Purpose: Download, EDA, and BIO NER transformation of the ai4privacy/pii-masking-200k dataset → GCS
• Status: ✅ Active
• Tasks: download_from_huggingface → load_raw_data → perform_eda → transform_data → upload_processed_data

Multi-Agent Debate (MAD) Output Guardrail

• Location: multi_agent/ (core pipeline), multi_agent_debate/ (with SQLite storage)
• Purpose: Verifies enterprise LLM answers against a regulatory evidence corpus. Extracts atomic claims, runs a 2-cycle adversarial debate (Agent A vs Agent B), routes through a partially-blind judge, and returns a routing decision (DELIVER / RETRY / HARD_BLOCK / HUMAN_REVIEW)
• Status: ✅ Active
• Run: uvicorn multi_agent.api:app --port 8001 --reload

🔮 Planned / In Development

Gateway — Input Guardrail

• Location: gateway/
• Purpose: 3 parallel classifiers (prompt injection, PII detection, jailbreak detection) with a weighted decision engine. Blocks threats before they reach the LLM.
• Status: 🚧 In Development

RAG Pipeline

• Location: rag/
• Purpose: Qdrant vector store + Qwen3-4B embedder + BM25 hybrid + cross-encoder reranker over a 72+ regulatory document corpus
• Status: 🚧 In Development

Confidence Scoring Engine

• Location: confidence/
• Purpose: Compute a final confidence score from LLM faithfulness, hallucination rate, RAG relevancy, and judge evaluation
• Status: 🚧 In Development

GRPO Feedback Loop

• Location: rlhf/
• Purpose: GRPO-based fine-tuning loop for Agent A (Brier reward) and Agent B (precision reward) using data written by the MAD pipeline to SQLite
• Status: 🚧 In Development
• Deployment handoff: GRPO-finetuned Agent A/B LoRA adapters are shared through Drive. See GRPO LoRA Deployment Handoff for adapter file manifest, vLLM model names (agent_a, agent_b), prompts, JSON-only schema, generation settings, AWS serving notes, and held-out eval results.

Finetuning Pipelines

• Location: finetuning/
• Purpose: Fine-tuning scripts for RoBERTa (jailbreak + PII), Llama-Prompt-Guard-2-86M (prompt injection), Qwen2.5-3B-Instruct (LLM generator), and Qwen3-4B (RAG embedder)
• Status: 🚧 In Development

Synthetic Evaluation Dataset

• Location: synthetic_data/
• Purpose: Generate 200-example healthcare-domain evaluation set (4 error types: fully_correct, missing_caveat, hallucinated_specific, jurisdiction_blind) for MAD pipeline benchmarking
• Status: 🚧 In Development

🛠️ Development

Prerequisites

• Docker & Docker Compose
• Python 3.10+
• Google Cloud Platform account (for GCS)
• Ollama (for MAD pipeline — ollama pull qwen2.5:7b)

Airflow Setup

# Setup Airflow
./scripts/setup_airflow.sh

# Start services
./scripts/start_airflow.sh

# Stop services
./scripts/stop_airflow.sh

MAD Pipeline (Quick Start)

pip install -r multi_agent/requirements.txt
ollama pull qwen2.5:7b && ollama serve
python -m multi_agent.run_test

📚 Documentation

• PII NER Pipeline Docs
• Quick Start Guide
• Setup Guide
• MAD Setup Guide