React client
Captures customer text and displays the category with a visual confidence score.
AI engineering case study
Every message,
understood.
A full-stack classification service that turns unstructured customer text into a reliable category and confidence score through a clean REST API.
4CategoriesRESTIntegration0-1Confidence3/3Tests passing
01 / The challenge
Customer messages arrive without structure.
Support teams receive complaints, questions, feedback and unrelated messages in the same stream. Manual triage slows response time and creates inconsistent routing. This project provides a small, integration-ready service that applies one shared classification policy to every message.
02 / Architecture
A focused path from input to insight.
Express API
Validates requests and exposes a focused POST /api/classify endpoint.
Ollama model
Uses llama3.2:3b with a deterministic prompt and strict JSON output.
MongoDB
Optionally stores successful classifications through Mongoose.
POST /api/classifyJSON
{
"text": "My order arrived damaged."
}
{
"category": "Complaint",
"confidence": 0.96
}03 / API design
One endpoint.
Predictable output.
The API accepts a text string, validates it, invokes the model service and normalizes the response. Consumers always receive an allowed category and a confidence value between zero and one.
Check API health04 / Engineering decisions
AI output treated as untrusted input.
Allow-listed categories
Model output is matched case-insensitively against Complaint, Query, Feedback and Other.
Bounded confidence
Numeric confidence is clamped to the valid 0-1 range, with a documented fallback.
Defensive validation
Empty text and payloads over 5,000 characters return clear HTTP 400 responses.
Optional persistence
The classifier remains available when MongoDB is not configured or connected.
05 / Technology
A stack where every tool has a clear responsibility.
The architecture for AI Text Classifier was shaped around the product workflow, integration boundaries and long-term maintainability. Technologies are used for specific responsibilities rather than added only to make the stack appear larger.
React
Builds reusable, state-driven interfaces and keeps complex user journeys organized into maintainable components.
Vite
Provides a fast development server and optimized production bundling for the client application.
Node.js
Runs the server-side JavaScript services, integration logic and non-blocking application workflows.
Express
Defines focused HTTP routes, middleware, validation and consistent REST API responses.
MongoDB
Stores flexible application records and supports evolving product data without rigid table structures.
Mongoose
Adds schemas, validation and model-level consistency to MongoDB persistence workflows.
Ollama
Runs the language model used for local, controlled AI inference and structured classification output.
LangChain
Organizes prompts and model calls behind a reusable service layer that can evolve independently.
Vercel
Hosts the web experience and serverless endpoints with repeatable deployment from source control.
Cloudflare Tunnel
Creates a protected route from the deployed API to the Ollama runtime without exposing local network ports.
Clear boundaries
Interface, business logic, persistence and external services remain separated so changes are easier to reason about.
Defensive data handling
Validation and predictable contracts protect downstream workflows from incomplete or unexpected input.
Deployment-aware design
Runtime limits, secrets, external dependencies and failure paths are considered as part of the implementation.
06 / Delivered
Built, tested and documented.
Responsive React testing interface
Layered controller, route and service structure
Automated API and normalization tests
Postman collection and environment examples
Vercel serverless deployment configuration
Detailed local and production setup guide
Explore the working product