Skip to content
NR
06/Case Study

AI Expense Management

NLP Categorization & Anomaly Detection

RoleFull-Stack AI Engineer
Timeline2025
StackOpenAI API, FastAPI, React
StatusShipped
Impact

89% auto-categorization accuracy with AI-driven spending anomaly detection and custom dashboards.

Overview

Context

This project demonstrates full-stack AI product engineering — combining NLP auto-categorization with financial anomaly detection, wrapped in a polished React frontend with interactive data visualizations.

Challenge

The problem

Manual expense categorization is tedious and error-prone. Users need intelligent auto-categorization they can trust, combined with proactive anomaly detection that surfaces unusual spending patterns without generating false alarms.

Approach

How I built it

01

Built NLP auto-categorization engine using OpenAI API with confidence scoring and manual override UX

02

Implemented AI-driven spending anomaly detection for proactive alerts

03

Created custom Chart.js dashboards for spending visualization and trend analysis

04

Built React + Redux frontend with TypeScript for type-safe development

05

Designed FastAPI backend with PostgreSQL, JWT auth, and Docker Compose deployment

Technical Decisions

Why these choices

Confidence scoring with manual override

AI categorization isn't always right. Showing confidence scores lets users trust high-confidence results and correct low-confidence ones — building trust through transparency.

Chart.js for data visualization

Lightweight, highly customizable, and well-suited for financial dashboards without the overhead of heavier visualization libraries.

Outcomes

What shipped

89% auto-categorization accuracy
AI-driven anomaly detection for spending patterns
Interactive Chart.js dashboards
Full-stack deployment on AWS EC2 with Nginx and CI/CD
Takeaways

What I learned

AI confidence scoring transforms user trust — transparency about uncertainty is a feature, not a weakness
Financial AI products require especially careful error handling — wrong categorizations erode trust fast