Case Study: PiggyVest Support Triage Agent

Documenting my process of building a firstline support triage agent for PiggyVest that is AI-powered. I am a fullstack developer who has been building software products for over 9 years now. My main focus has been server and UI (mobile and web) development mostly using Typescript (Node.js, React, React Native). I have a goal to learn about AI product development, that is what drives me to work on this case study. Also I see this as the next frontier in software development and I want to be a part of it.

Background

PiggyVest is a Nigerian fintech company that provides a platform for users to save and invest their money.

Problem

I will be using real, messy data instead of a synthetic dataset: manually screenshotted PiggyVest Twitter replies, OCR'd with Gemini. We will also be using publicly available FAQs, blogs and other relevant information to provide context for the AI agent.

This is purely for educational purposes (I hope this is not a violation of their terms of service but this is for personal use and not for commercial use).

Why this project?

This project is a good and common use case for AI agents. A support triage agent has a clear success metric (accuracy against human-labeled ground truth), real user issues and a clear learning progression (gathering data -> cleaning data -> calling an LLM -> structured output -> evaluation -> improvement).

What's ahead

• Building the support triage agent platform
  • Backend built with Node.js and PostgreSQL
  • Frontend built with React (Vite)
  • Dockerized
• Binary check evaluation
• LLM-as-judge evaluation
• RAG (Retrieval-Augmented Generation)
• Tool calling
• Function calling
• Agent response generation/auto escalation

Final Goal

The aim of this project is personal development and learning. I want to know what it takes to build a full functional AI agent.

And of course see how my skills in software development can translate into AI product development.