SpendWise: AI-Powered Personal Finance with Structured Validation

Overview

SpendWise is an AI-powered personal finance platform that helps users track income, expenses, and budgets while providing intelligent spending insights. The core technical challenge was building a reliable AI pipeline using Google Gemini that produces structured, deterministic financial data from a non-deterministic language model.

Problem

Personal finance apps need accurate categorization to be useful, but AI models produce free-form natural language that is difficult to parse into structured financial data. Off-the-shelf solutions either use rigid rule-based systems that miss nuances or pass raw AI output straight to the user without validation.

Solution

Implemented a two-stage pipeline: a categorization stage that sends transaction data to Gemini with predefined spending buckets, followed by a validation stage that parses the output, checks against an allowed category list, and rejects anything ambiguous with a retry using a more constrained prompt.

Technical Architecture

• Frontend: Next.js 14 App Router, Tailwind CSS, Framer Motion
• Backend: Next.js API routes, Prisma ORM, MongoDB
• Auth: NextAuth.js with Google OAuth + credentials
• AI: Google Gemini for spending insights and categorization
• Caching: Redis for dashboard data and session management

Results

• 100% AI response validation rate after implementing the two-stage pipeline
• Budget snapshot system preserves historical data even after edits
• Sub-second dashboard loads via Redis cache layer
• Two-factor authentication for sensitive operations