CARE (Customer Assistance and Response Engine) is an advanced AI-powered chatbot designed to provide efficient, empathetic, and domain-specific support across various industries, including:
- Banking
- Healthcare
- E-commerce
- Telecommunications
The chatbot leverages Phi 3.5 Mini Instruct, a compact, instruction-tuned language model, optimized for efficient fine-tuning and deployment. CARE aims to automate customer interactions, analyze sentiment, and escalate unresolved issues seamlessly.
- Domain-Specific Responses: Customized support for diverse industries like healthcare, banking, telecommunications, and e-commerce.
- Sentiment Analysis: Provides empathetic responses by analyzing user sentiment.
- Efficient Performance: Built using a compact model optimized for faster inference and lower memory requirements.
- Extensible Framework: Easily expandable to support additional industries or integrate with new systems.
The project employs the following workflow:
- Input: User prompts specific to the industry (e.g., banking, healthcare).
- Processing:
- Sentiment analysis for empathetic responses.
- Fine-tuned language model tailored for the respective domain.
- Output: Generates accurate, domain-specific responses and escalates issues if needed.
Phi 3.5 Mini Instruct ensures:
- Instruction-tuned training for contextual understanding.
- Transformer architecture for balanced performance and efficiency.
CARE is trained on domain-specific datasets to enhance accuracy:
- Banking: Consumer Complaints Dataset
- Healthcare: AI Medical Chatbot Dataset
- Telecommunications: Telco LLM Chatbot Training Dataset
- E-commerce: E-commerce FAQs
Custom datasets were also utilized for enhanced domain adaptation.
- Model Architecture: Fine-tuning and QLoRA techniques were used to optimize Phi 3.5 Mini Instruct.
- Training Pipeline:
- Data preprocessing: Addressed dataset skewness using oversampling and undersampling techniques.
- Model merging: Combined fine-tuned domain-specific models into a unified framework.
- Final Model: Phi-3.5-mini-instruct-Combined-Model
- Testing: Testing Notebook
Future enhancements for CARE could include:
- Multi-Language Support: Expand capabilities to handle queries in multiple languages.
- Voice Integration: Real-time voice-to-text and text-to-voice for improved accessibility.
- Reinforcement Learning with Human Feedback (RLHF): Continuously improve chatbot accuracy using RLHF and Retrieval-Augmented Generation (RAG).
- Knowledge Base Expansion: Regularly update databases for better accuracy and relevance.
- API Deployment: Make the chatbot available as an API to expand accessibility and impact.
CARE represents a step forward in smart customer service, providing solutions tailored to diverse industries. By leveraging advanced NLP techniques and domain-specific training, CARE enhances customer satisfaction and streamlines issue resolution.
- Ankit Dutta
- Nabarup Ghosh
- Ankush Chatterjee
- Abhishek Das
Special thanks to Genpact and Kshitij for providing the opportunity to develop this impactful project.
This project is licensed under the AGPL-3.0 License. See the LICENSE file for details.