Professional Data Science Portfolio

---

Improved Search for Text Rich Knowledge Graphs

In this research proposal, we aim to enhance the ranking of entities within knowledge graphs used in the healthcare sector by leveraging advanced natural language processing techniques. Through a comprehensive comparative study, we will explore the effectiveness of both graph-based ranking and embedding-based ranking approaches. Our proposed architecture, which involves data extraction, integration, and machine learning-based reasoning, ensures scalability, reliability, and interoperability, thereby enabling better decision-making, trend identification, and improved patient care in pharmaceutical companies and healthcare organizations.

Research Work

---

Enabling Continuous Improvement in Machine Learning Workflows through Data Logging and Monitoring

The complexity of deploying data-intensive machine learning models in real-world applications requires effective MLOps practices. This thesis proposes an open-source solution for data logging and monitoring using tools like MLFlow and WhyLogs. It also addresses model drift with techniques to detect and mitigate it, ensuring model accuracy and reliability.

Business Usecase Presentation • Thesis Documentation

---

Low-Cost Image-based Diagnosis System for Cleft Palate Reconstruction using 3D-Neural Modelling.

Palatoplasty, or the surgical treatment of children born with cleft palate, is continuously evolving, and there are several preoperative issues that a cleft surgeon needs to consider before embarking on a palatal surgery. Hence, cleft palate surgery remains an enigma for most cleft surgeons. The proposal aims to cater to millions of underprivileged infants and adults by lowering the treatment cost in the diagnosis and pre-surgery therapy. This work proposes a 3D imaging technology based on deep learning that provides a comprehensive recording of the facial morphology that lends itself to the objective and subjective assessment of cleft palate repair surgery.

Research Work

---

Conversational AI Chatbot Using Reinforcement Learning

This NLU Chatbot project aimed to build a scalable, self-learning conversational AI chatbot using reinforcement learning. The model was trained on a dataset of large fictional conversations extracted from movies, achieving a BLEU score of 0.86. The approach used self-critical sequence training to optimize multiple target sequencing and improve convergence. The saved model was deployed to Telegram as a DBS-Assistant chatbot using Heroku, capable of context-based basic conversation and retaining memory. The project was implemented with a RNN and LSTM backbone, and trained with a batch size of 32, learning rate 1e-3 for 100 epochs.

Business Usecase • Project Documentation

---

Gender Identification of Bloggers for Personalized Content Recommendation

This project aims to identify and detect the gender of bloggers based on textual data to personalize content recommendation. The dataset comprises ~2,599 records with 1,300 females and 1,299 males. After preprocessing and exploring the data, 7 different machine learning algorithms were trained, and Naïve Bayes and Multi-Layer Perceptron were the best performing models with accuracy rates of 70% and 72%, respectively. In RapidMiner, the Support Vector Machine model had the least classification error of 32.7%, while Decision Tree was faster but with a higher classification error. As a result, fine-tuned machine learning models outperform RapidMiner metrics.

View Jupyter Notebook • Project Documentation

---

Experience

2022

Business Intelligence Engineer - AspireNXT Pvt. Ltd.

2020

Research Data Scientist - ZeoMinds Pvt. Ltd.

2019

Data Analyst - Digital Nest Pvt. Ltd.

2018

Android Developer - E Info Solutions Pvt. Ltd.

---

Contact Me