An AI-Powered On-Device Expense Tracking System Using BiLSTM-Based Named Entity Recognition on Financial SMS Notifications

Abstract

The proliferation of digital payment platforms in India—including UPI-based services such as Google Pay, PhonePe, and Paytm—has led to a dramatic increase in transactional SMS and push notification volume, creating an opportunity for automated personal finance tracking. This paper presents Notitrack, an intelligent, offline-first mobile expense tracking system that leverages a custom-trained Bidirectional Long Short-Term Memory (BiLSTM) neural network for Named Entity Recognition (NER) to automatically extract financial entities—specifically merchant names, transaction amounts, and bank identifiers—from banking SMS messages and push notifications. The system is implemented as a cross-platform Flutter application with Android-native notification interception via the NotificationListenerService API. A hybrid extraction architecture is employed, combining on-device TensorFlow Lite inference with a multi-tier regex fallback system to ensure robust entity extraction even with out-of-vocabulary tokens. The BiLSTM model was trained on a custom-annotated dataset of 211 Indian banking SMS samples using IOB (Inside-Outside-Beginning) tagging, focal loss for class imbalance, and iteratively tuned per-tag class weighting. Experimental results demonstrate an overall tagging accuracy of 95.03%, with entity-specific F1-scores of 0.97 for transaction amounts, 0.97 for bank names, and 0.69 for merchant names. The system processes transactions in under 100 milliseconds, operates entirely offline with zero cloud dependency, and includes a two-layer duplicate prevention mechanism. This work contributes a practical, privacy-preserving approach to automated expense management on resource-constrained mobile devices.