Client: Plus One Company
Role: NLP Engineer | ML Engineer
Project: NLP-Powered Call Analysis Platform
Developed and deployed a comprehensive NLP-powered call analysis platform that supports clients, agents, campaigns, and sales interactions. The platform leverages advanced machine learning techniques to extract insights from voice calls and enable data-driven decision-making.
Technologies and Tools:
- Programming Language: Python
- NLP Libraries: spaCy, scikit-learn, BERT
- Google Cloud Platform (GCP) Services:
- BigQuery, Cloud SQL, App Engine, Cloud Storage, Compute Engine, Cloud Functions, Cloud Memory Store, Cloud Scheduler, Cloud Run, Kubernetes Engine, Artifact Registry, Source Repositories, PubSub, AI Platform, AI Notebook, Cloud Natural Language, AutoML Tables, Cloud AutoML
Key Accomplishments:
- Developed an NLP worker engine that extracts key phrases from thousands of voice calls daily using unsupervised machine learning techniques.
- Utilized Google Speech-to-Text API to convert voice calls into textual data.
- Implemented and optimized unsupervised keyphrase extraction algorithms, including Topicrank, Textrank, Yake, and Autophrase.
- Employed ensembles and various pre and post-processing techniques to aggregate top keyphrases.
- Deployed the NLP worker on Google App Engine for scalable and reliable processing of voice calls.
- Optimized the performance of the NLP worker by leveraging Google Cloud Memory Store.
- Ensured efficient resource utilization and seamless integration with other GCP services.
- Utilized Google Cloud AutoML and AutoML Tables to build keyphrase and call segmentation models on human-annotated data.
- Trained, evaluated, and served models using AutoML, automating the model development process.
- Implemented automated model retraining to adapt to evolving data patterns and maintain high accuracy.
- Performed sentiment analysis using Google Natural Language API and VADER (Valence Aware Dictionary and sEntiment Reasoner).
- Integrated sentiment scores generated by these algorithms as features for training a custom sentiment analysis model using BERT.
- Achieved improved sentiment classification accuracy by leveraging the power of transfer learning with BERT.
- Researched, prototyped, and optimized state-of-the-art machine learning and deep learning techniques for student text classification.
- Implemented and fine-tuned models such as SVM, Logistic Regression, Random Forest regression, LSTM, and CNN using scikit-learn, Keras, and TensorFlow.
- Conducted extensive hyperparameter tuning to optimize model performance on CPU/GPU environments.
- Employed model interpretation techniques, including ELI5 (Explain Like I'm 5) and LIME (Local Interpretable Model-Agnostic Explanations), to provide insights into model predictions.
- Analyzed feature importances and local explanations to understand the factors influencing model decisions.
- Utilized these insights to refine feature engineering and improve model interpretability.
- Developed an ETL pipeline to ingest data from Cloud SQL to BigQuery, enabling scalable data processing and analysis.
- Designed and implemented an ETL pipeline capable of handling terabytes of data and running continuously throughout the day.
- Optimized data transformations and joins using complex SQL queries in BigQuery for efficient data preparation.
- Created interactive dashboards using Google Data Studio by connecting to BigQuery for analyzing keyphrase data.
- Designed visually appealing and informative dashboards to present key insights and trends.
- Enabled stakeholders to explore and derive actionable insights from the extracted keyphrase data.
Key Achievement: