AyushAudit AI — The Story

Inspiration

India's PM-JAY insures 500M+ people yet loses an estimated ₹2,000 Cr/year to fraud. A rule-based auditor felt solvable with RAG.

How it's built

$$\text{Audit} = \text{RAG}(\text{Guidelines}) \xrightarrow{\text{LLM}} P(\text{fraud}) \gtrless \tau$$

Guidelines → chunked → FAISS index → top-$k$ rules retrieved per claim → LLM judges.

What I learned

• Unity Catalog replaced DBFS silently — every path assumption broke
• hive_metastore is deprecated; /tmp is ephemeral; Gradio beats Streamlit inside notebooks
• Heuristics are a 必须 fallback when APIs fail

Challenges faced

The real lesson

$$\text{Production} = \text{Good idea} + \underbrace{\text{10x debugging}}_{\text{the actual work}}$$

Built for the 55M PM-JAY claims processed annually — catching fraud before reimbursement, not after.

Built With

• datalake
• python
• replit