NirviSha: AI Financial Agent for CAs

NirviSha is a AI agent designed to solve the manual data entry for CAs. It uses Gemini 3's multimodal feature to read receipts and POs and instantly digitize them into a structured database.

Comment

Inspiration

In India, Chartered Accountants (CAs) and tax professionals face a chaotic monthly ritual: collecting thousands of financial documents from clients. These documents arrive as crumpled receipts, blurry WhatsApp photos, or PDFs scattered across emails.

We realized that while CAs use sophisticated software for filing taxes, the process of collecting and digitizing the data is still manually done by junior staff. We asked: "What if a CA's client could just WhatsApp a photo of their bill, and an AI Agent instantly structured it into the accounting ledger?"

That's how NirviSha was born.

What it does

NirviSha is a Multi-Tenant SaaS Platform integrated with a WhatsApp AI Agent.

  1. For the Client: They simply chat with our verified WhatsApp Bot. They send images of invoices, bills, or petrol receipts.
  2. The AI Brain (Gemini 3): The bot uses Google Gemini's multimodal capabilities to "see" the image. It extracts key fields: Date, Merchant, Amount, Tax, and Invoice Number.
  3. For the CA: The data appears instantly on their Web Dashboard, categorized and ready for review. No data entry required.

How we built it

We built NirviSha using a modern, scalable stack:

  • Google Gemini API: We utilized the latest Gemini Vision models to handle the OCR and entity extraction. We chose Gemini for its superior ability to understand Indian context (e.g., distinguishing between "SGST" and "Total").
  • WhatsApp Cloud API: For the frictionless user interface that everyone in India already knows how to use.
  • Next.js (React): For the robust, multi-tenant web dashboard.
  • Supabase (PostgreSQL): For real-time database storage and strict Row Level Security (RLS) to ensure data privacy between different CA firms.
  • Vercel: For serverless deployment and webhook management.

Challenges we faced

  • Meta Verification: Navigating the strict WhatsApp Business API verification process to get our bot online was a significant hurdle.
  • Data Privacy: Since we are dealing with financial data, we had to architect a strict "Tenant Isolation" system using Supabase RLS so one CA firm could never see another's data.
  • Unstructured Inputs: Indian receipts are notoriously non-standard (handwritten totals, faded ink). Tuning the Gemini prompt to handle these edge cases was a key technical challenge.

What we learned

We learned that "Chat is the new UI." By moving the complex data entry task into a simple WhatsApp chat, we drastically reduced the friction for the end-user. We also learned the immense power of Gemini's Multimodality—it didn't just read the text; it understood the context of the document (e.g., knowing that a "Starbucks" receipt is likely "Food & Beverage").

What's next for NirviSha

We plan to introduce "anomaly detection" where Gemini warns the user if a bill looks like a duplicate or if the tax amount doesn't match the percentage, effectively acting as a first-line auditor.

Built With

Share this project:

Updates