Hisab AI Agent

Hisab Kitab already helped users manage expenses, accounts, contacts, and udhar, but adding records manually could feel slow. Users often think in casual language like “NayaPay se 420 ka dinner add kar do” instead of structured forms.

The challenge was to make this feel conversational while still keeping financial actions safe. The AI should understand the request, but it should not silently save anything important without the user confirming it first.

I built a FastAPI-based AI agent that connects the Hisab Kitab app with OpenAI and the existing Hisab API.

The agent can understand Roman Urdu and English messages, fetch user-specific data, propose expenses, settlements, and udhar entries, and return a pending action card for the frontend to show.

The key safety feature is the propose-confirm flow: the AI prepares the action, but the user must click Confirm before anything is saved. I also added persistent chat sessions, pending action storage, per-user scoping, protected API routes, run logging, and eval tests to measure accuracy.

The result is a much smoother way to use Hisab Kitab. Instead of filling multiple fields manually, a user can write a natural sentence and review the proposed result before saving.

It also made the AI safer for real financial data. The agent does not directly write records on its own, and each user’s chat sessions and pending actions are kept separate.

In eval testing, the agent improved from 75% to 100% across 20 real test cases covering expenses, settlements, udhar, clarification, and refusal behavior.