TextUs

Jan-Apr 2025

TextUs is an AI-powered customer training platform, helping customer representative associates learn how to handle real-world customer queries by simulating conversations with AI agents.

Overview

We designed TextUs after learning about the challenges of training customer service representatives from CPF Board, a Singaporean government agency.

As customer service shifts from conventional inquiry-based systems to live chats, we dug deep with CPF's staff to understand how tech can transform their processes.

The big question: how do we get new customer service associates ready to tackle complex convos with accuracy, speed, and confidence when things get crazy?

Contributions

• Designed and architected a full-stack AI training simulator that lets Customer Service Officers practice simultaneous chats with multiple AI-driven “Scenario Customers”, improving response accuracy and speed.
• Built Retrieval-Augmented Generation (RAG) pipelines that enrich GPT-4o responses with CPF policy data, boosting factual accuracy and reducing hallucinations during trainee interactions.
• Developed a FastAPI backend with async SQLModel and PostgreSQL, delivering JWT-secured and rate-limited APIs.
• Engineered a responsive Next.js + Tailwind front-end with live WebSocket updates
• Introduced structured rubric evaluation: created database models, REST endpoints, and UI that let trainers score sessions on tone, comprehension, and accuracy, and feed results back into trainee dashboards
• Added many-to-many file-upload support - reference documents, rubrics, and scenario assets - and wired it end-to-end across database, API, and React components.

Key Features

1. LangGraph conversation engine: A stateful workflow system where each interaction is a directed graph with checkpoint-based resumption.
2. Multi-agent simulation: Multiple AI customers bombard associates simultaneously, each with their own quirky personalities, backgrounds, patience levels with realistic response delays based on temperament and message length - all to simulate real-world customer interactions.
3. Structured rubrics: Post-session scoring and feedback for managers to track associates' progress with precision.
4. RAG knowledge base: Modular and keeps our AI queries and evaluations in sync with the most updated info.

Implementation

• Frontend: Next.js, TailwindCSS, Shadcn UI
• Backend: FastAPI, LangGraph
• LLM: OpenAI GPT-4o, Google Gemini embeddings
• Hosting: Docker, on Render

Useful Links

I share my learnings from this project in a blog post.

GitHub Repo: https://github.com/pyraxo/textus