An NLP system designed to classify support tickets automatically and reduce manual routing work.
The challenge
Manual ticket sorting can be slow and inconsistent, making prioritization and response management difficult.
The strategy
Build a reproducible text-classification pipeline, compare it with a simple majority baseline, and expose predictions through a user-friendly application.
My contribution and execution
- Cleaned and prepared noisy ticket text.
- Explored category distributions and text patterns.
- Built a TF-IDF and logistic-regression baseline and designed a stronger contextual modeling path.
- Created deployment architecture, a prediction interface and a CI/CD workflow.
Tools and methods
Python, NLP, TF-IDF, logistic regression, transformer experimentation, FastAPI, Gradio, Docker and CI/CD
Outcome, evidence and limitations
The project report records a 19.2% majority-class baseline and approximately 95% accuracy for the TF-IDF logistic-regression model on its evaluation split. Production readiness would still require class-level metrics, leakage checks, drift monitoring and testing on unseen operational data.
Why this project matters
This project demonstrates how I translate a practical problem into a structured strategy, prototype or analytical workflow. It also records the limits of the current work and the next validation needed before broader use.
← Back to all insights