OCBC Bank - AI-Powered Transaction & Sentiment Analysis System
1. Project Overview
This project aims to develop a web and mobile application that leverages Large
Language Models (LLMs) and real-time data visualization to clean and
analyze OCBC’s credit card transactions and customer feedback.
The solution will:
Automate cleaning of messy/unstructured transaction data.
Identify sentiment trends in customer feedback.
Generate real-time data insights for internal OCBC users (IT & Marketing).
Provide an intuitive, user-friendly interface for seamless interaction.
Ensure scalability, accuracy, and efficiency in processing large datasets.
2. System Features & Functionalities
A. Data Processing & Cleaning
Accepts multiple file formats (CSV, JSON, Excel).
Validates file type before processing.
Classifies as either Customer Feedback or Credit Card Transactions.
Cleans and standardizes messy transaction data using LLM-driven NLP
techniques.
Detects spam/irrelevant customer feedback and filters out noise.
B. Sentiment Analysis & Real-Time Insights
Uses DeepSeek / Llama 3.1:3B for high-accuracy analysis.
Option to switch to Paraphrase-MPNet-Base-V2 for faster but less
accurate insights.
Generates real-time sentiment tracking with Power BI / Tableau /
[Link].
C. Real-Time Data Visualization
Live dashboard to track sentiment trends.
Color-coded indicators for negative spikes (e.g., red for major
complaints).
Historical data trends for long-term insights.
Real-time alerting on major sentiment shifts.
D. User Interface
1. Web Application
Developed in [Link] with a [Link] & FastAPI backend.
Intuitive interface with a hamburger menu for:
o Customer Feedback Analysis
o Credit Card Transaction Cleaning
Interactive data filtering & export functionality.
2. Mobile Mini-App (Android Studio)
Streamlined interface for sentiment tracking on the go.
Ability to upload and analyze customer feedback via mobile.
E. Backend & Infrastructure
Database: MongoDB (scalable, handles large unstructured datasets).
Queue System: Redis / RabbitMQ to manage multiple file uploads.
Pre-processing pipeline to optimize LLM response time.
3. Technical Architecture
Frontend: [Link] (Web), Kotlin (Android Mini-App)
Backend: FastAPI (Python) & [Link] for API integration
Database: MongoDB for high-speed data retrieval
ML Model Hosting: Hugging Face / AWS Sagemaker
Visualization: Power BI / Tableau / [Link]
Queue Management: Redis / RabbitMQ
LLM Models: DeepSeek / Llama 3.1:3B + Paraphrase-MPNet-Base-V2
4. Project Phases
Phase Tasks
Sandbox Model training, accuracy testing, data validation
Pre- UI design, backend integration, performance
Production optimization
Full-scale deployment, internal testing, future
Production
enhancements
5. Challenges Faced
1. Accuracy vs. Speed Trade-off
o LLMs can be slow, requiring optimizations.
o Used distilled models for faster processing in some cases.
2. Real-Time Data Processing Bottlenecks
o Optimized pipeline with batch processing & queueing.
3. Model Training Challenges
o Faced data scarcity issues, overcame by generating synthetic
feedback data.
4. Scalability Concerns
o Ensured MongoDB indexing & Redis caching for smooth
performance.
5. User Adoption & UI Design
o Iterated UI based on feedback for better usability.
6. Future Enhancements
Advanced personalization per OCBC department’s needs.
Multi-language support for regional teams.
Integration with OCBC’s enterprise analytics tools.
Production-scale optimization & load testing.
7. References (APA Style)
Brown, T. et al. (2020). “Language Models Are Few-Shot Learners.”
NeurIPS.
Devlin, J., et al. (2018). “BERT: Pre-training of Deep Bidirectional
Transformers for Language Understanding.” ACL.
Microsoft. (2021). “Paraphrase-MPNet-Base-V2: Sentence Embedding
Model.” Hugging Face Models.