Swahili SMS Spam Detection.
Swahili SMS Spam Detection System
This project focuses on detecting Mobile Money scams in East African SMS messages that combine Swahili with informal or misspelled English — a linguistic blend common in real-world communications.
🔍 Problem Statement
SMS scams in East Africa often target victims via Mobile Money fraud. These messages are not only written in Swahili but frequently include English words with informal or misspelled phrasing, making traditional spam classifiers less effective.
💡 Solution Overview
- Built a spam detection system using BERT-based embeddings to understand the semantic context of mixed-language messages.
- Developed a FastAPI web service for serving predictions in real time.
- Leveraged LLMs including GPT, LLaMA, Gemma, and Qwen for:
- Zero-shot classification
- Model explainability
- Topic modeling to uncover common scam themes
- Implemented MLOps with DVC to version control datasets, models, and experiments.
- Tracked experiments with MLFlow.
🛠️ Tech Stack
- Python, FastAPI
- BERT (LaBSE)
- GPT, LLaMA, Gemma, Qwen
- DVC for experiment tracking and reproducibility
🚀 Key Outcomes
- Successfully detected scam patterns in highly code-mixed text messages
- Demonstrated zero-shot capabilities using multiple LLMs for both classification and interpretability
- Ensured reproducibility with end-to-end MLOps practices
