ComplianceGuard RAG System

Inspiration

Many companies struggle to ensure product features comply with geo-specific regulations, especially when teams rely on generic LLMs that misinterpret domain-specific jargon or hallucinate answers. We wanted to create a system that augments AI with regulatory knowledge to help teams detect compliance risks quickly and reliably.

What it does

ComplianceGuard automatically flags product features that may require geo-specific compliance logic. By combining retrieval-augmented generation (RAG) with a fine-tuned LLaMA 3.2 1B model, it:

• Understands internal feature codenames and abbreviations
  
• Retrieves relevant regulatory context from a vector database
  
• Generates transparent, audit-ready reasoning with references to regulations
  

How we built it

• Data Collection & Knowledge Base: Compiled regulations like DSA and California Kids Act into a structured JSON database.
  
• Embedding & Retrieval: Used SentenceTransformers to generate embeddings and ChromaDB for fast semantic search.
  
• LLM & Fine-Tuning: Leveraged Meta LLaMA 3.2 1B Instruct, fine-tuned with LoRA to adapt the model to compliance-specific reasoning.
  
• Pipeline: Integrated embeddings retrieval with LLaMA reasoning for accurate, context-aware compliance assessments.
  

Challenges we ran into

• AI Hallucinations: Initial model often misclassified features due to vague prompts.
  
• Domain-Specific Jargon: Mapping internal feature codes to regulatory concepts required careful preprocessing.
  
• Notebook & Metadata Issues: Encountered rendering problems on GitHub with .ipynb files due to widget metadata.
  

Accomplishments that we're proud of

• Built a fully functional RAG pipeline that integrates embeddings and LLaMA reasoning.
  
• Successfully fine-tuned LLaMA with LoRA for domain-specific compliance tasks.
  
• Generated audit-ready outputs with clear, regulation-backed reasoning.
  

What we learned

• Retrieval-augmented generation dramatically improves LLM reliability for niche domains.
  
• Fine-tuning with LoRA allows adaptation to specialized tasks

Built With

• chromadb
• langchain
• llama
• peft
• pydantic
• python
• sentence-transformers
• torch
• transofrmers