VIVIRITY Intelli-Credit

Credit risk intelligence system built on a non-embedding RAG pipeline. 2nd place at the IIT Hyderabad AI/ML Hackathon (YUVAAN 2026).

status completed

stack Python, FastAPI, Google Generative AI Toolkit, Google A2A Toolkit, Gemini API, RAG, Pandas, NumPy

code https://github.com/charansaiponnada/VIVIRITY

An end-to-end AI credit intelligence system that reads annual reports and turns them into structured loan risk assessments. Built for YUVAAN 2026 at IIT Hyderabad, where it placed 2nd among the Top 10 finalists out of 7,600+ registrants.

problem

Credit analysts read 500+ page annual reports by hand to extract the handful of numbers that actually drive a lending decision. It takes days per document, and vector-based retrieval over that volume is expensive in API calls.

solution

A non-embedding RAG pipeline feeding a multi-agent decision pipeline, so retrieval happens without the cost of maintaining and querying an embedding index.

architecture

Ingestion: automated parsing of 500+ page annual reports
Retrieval: non-embedding RAG pipeline, no vector index to maintain
Extraction: 50+ financial risk indicators
Decisioning: multi-agent pipeline (Google A2A toolkit) for end-to-end loan risk assessment
API layer: FastAPI

results

98% lower API overhead than traditional vector-based retrieval, and extraction time down from multiple days of manual work to under 5 minutes. Placed 2nd among the Top 10 finalists at YUVAAN 2026, IIT Hyderabad, out of 7,600+ registrants.

questions

Why non-embedding retrieval?

Embedding every chunk of a 500-page report and querying a vector store was the dominant cost. Skipping the embedding step cut API overhead by 98% without losing the retrieval quality the decision pipeline needed.