Shubham Baghel

Project : Voice bot for medical domain

Goal : business objective is to reduce load from call center for appointment booking with doctors.

Solution:

• Build an AI bot utilizing capabilities of Speech to Text (STT), Rasa NLU and Text to Speech (TTS) modules.
• Collected medical domain specific data to extract entities and intent of user, integrated client booking API with bot. Deployed model as a service to integrate with telephony channel.

Result :

• Reduced 20-30% of workload from call center by building a voice bot to book an appointment with doctors.

Project : Aspect based Sentiment Analysis

Goal : business objective is to build a model to extract aspect and associate sentiment associated with customer interaction with agents

Solution:

• Fine tuned transformers based model on the multimodal task of intent classification and aspect extraction using PyABSA package using domain specific data.
• added translation in pipeline to support this model for vernacular language support.
• Deployed the solution as FastAPI. It can be integrated with multiple business.

Result :

• Help businesses to understand the common issues most of the customers are facing and overall sentiment towards product

Project : Question Answering system integration with Chabot

Goal : Build a Search system where student can ask any query related to academic.

Solution:

• Fetched Question and Answer data from database, transcripts of the content available.
• indexed questions and docs in ElasticSearch DB for retrieval, applied DPR algorithm for retrieve document based on contextual similarity.
• Fine tuned BERT model for QnA for academic domain.
• Deployed model as a flask API.

Result :

• Optimized the search for domain specific result for students.

Project : Recommendation Engine

Goal: Build a system that can recommend personalized content to students.

Solution:

• Requirements gathering from stockholders and academic team to generate different types of recommendation.
• Build a pipeline to extract data of student interaction with content and train the model or generate recommendation on updated data.
• Build apache airflow to create and schedule jobs for retraining of model.

Result :

• Generated AI based and schedule based recommendation to students.

Project : Extractive Q & A system from unstructured documents

Goal : Build a Question answering system to extract information from documents and search precisely within document.

Solution:

• Extracted data from PDFs and indexed paragraphs on ElasticSearch DB.
• Applied Sparse and Dense Passage Retrieval to find relevant paragraphs.
• Utilized BERT Q & A model fine tuned on SQUAD 2.0 dataset.

Result :

• Achieved 80% of f1 score on test data, provided effective search for users.

Project : Sentiment Extraction for customer behavior analysis in live chat environment

Goal : Build a AI model that can predict the sentiment in five labels

Solution:

• Collected open source data available for sentiment ea. reviews, conversations, web series.
• Applied custom LSTM based Neural Network architecture, due to less data applied transfer learning and fine-tuned AlBERT model.

Result :

• Achieved 85% of accuracy with five classes classification.
• Reduced inference time from 7 sec to few ms by trying different transformer based models.

• Developed a supply chain management system for crude oil transportation
• Integration of application for clients in same domain
• Taking client specific requirements and implementation of them in existing system

• Developed an ERP system for construction domain.