RBI-Artha-Sahayak

Inspiration

India has millions of first-time digital banking users who see error codes instead of explanations, and regulatory updates they can't understand. We wanted to fix that — give every person clarity over their own money, in their own language.

What it does

Artha Sahayak converts fraud alerts into plain-language messages, generates real-time financial health summaries, and translates RBI guideline updates into simple notifications — all delivered multilingually to users across Tier-2, Tier-3, and rural India.

How we built it

Built on Databricks Lakehouse with Delta Lake for data storage and Spark Structured Streaming for real-time transaction processing. An LLM layer fine-tuned on RBI documents generates the plain-language outputs, with a multilingual pipeline handling regional language delivery.

Challenges we ran into

Keeping AI explanations both simple and financially accurate was harder than expected. Translating banking terminology into regional languages without losing meaning was tricky. We also had to hit low-latency targets for real-time fraud alerts while keeping data handling compliant with RBI norms.

Accomplishments that we're proud of

Building a fraud explanation engine that produces calm, clear, multilingual alerts in real time — and an RBI guideline parser that turns dense circulars into one-paragraph user summaries within minutes of publication.

What we learned

Simplicity is the hardest engineering problem. We also learned that AI outputs which acknowledge uncertainty ("we think this may be...") build more user trust than overconfident ones — especially in fraud scenarios.

What's next for RBI-Artha-Sahayak

Expanding to all 22 scheduled Indian languages, adding a voice interface for non-smartphone users, building proactive financial coaching features, and opening the system as an API that any bank or fintech can integrate.

Built With

• databricks
• javascript
• metaflow
• ml
• rag
• vite