AI in Customer Service Automation for CRM Systems
A transformer-based conversational AI for customer support, with a comparative evaluation of rule-based, DistilBERT and Sentence-BERT approaches on accuracy, confidence and latency.
At a glance
- Type
- Bachelor’s thesis
- Period
- 2024 – 2025
- Technologies
- Python
- PyTorch
- DistilBERT
- Sentence-BERT
- Flask
- Java
- Spring Boot
- REST
- HTML / JS
Purpose
To apply DistilBERT and Sentence-BERT to customer-service automation for CRM systems, and to benchmark rule-based, DistilBERT and Sentence-BERT approaches on accuracy, confidence and latency in order to quantify the trade-off between interpretability and precision.
My contribution
I built a transformer-based conversational AI, using DistilBERT for intent classification and Sentence-BERT for semantic similarity retrieval, and benchmarked the rule-based, DistilBERT and Sentence-BERT approaches on accuracy, confidence and latency.
Key features
- Intent classification with a fine-tuned DistilBERT model.
- Semantic similarity retrieval with Sentence-BERT over a curated customer-support corpus.
- Two-service architecture: a Python / Flask ML inference micro-service and a Java / Spring Boot dialogue backend, plus an HTML / JavaScript front-end, using REST APIs.
- Comparative evaluation of rule-based, DistilBERT and Sentence-BERT approaches, benchmarking accuracy, confidence and latency.
System components
A schematic of the components named in my CV: a Python / Flask ML inference micro-service, a Java / Spring Boot dialogue backend and an HTML / JavaScript front-end, using REST APIs. It does not show how requests flow between them. It is a diagram, not a screenshot.