Rupali KaPatel

LLM Based Applications:

• Deficiency backfilling by generating Question using GPT+RAG pipeline
• Automatic lesson generation using gpt assistant
• Actionable Insights using chain of table approach

Doubt Resolution:

• Designed and led the development of a comprehensive system aimed at addressing user queries using doubt resolution bot.
• The system identifies contextually similar questions from a vast pool using state-of-the-art models fine-tuned with in-house academic corpora, which include question text and associated diagrams or figures.
• The solution was expanded to support 11 Indic languages using LABSE, enhancing B2G deals.

Smart Tagging

• Automatically assigning various meta tags to questions using fine-tuned MiniLM and XGBoost models, achieving a 70% reduction in manual labor.

Notifications: Led Notification team at Embibe.

• User Segmentation: Used K means clustering to analyze users across 15 dimensions to develop comprehensive user profiles for decision-making.
• Daily push notification based on student recent interaction with platform. LLM based notification generation, providing a diverse experience in terms of the notifications. 2x better than popularity based models.
• A matrix factorization based recommendation system aimed to deliver personalized content on homepage carousel. 1x better than popularity based models.

Insta-Solver

Solving natural language mathematical questions for K12 students with step-by-step explanations.

https://www.embibe.com/in-en/artificial-intelligence-ai-in-education-articles/instasolver/

Question Generation

A system built with the SOTA T5 model which enable students to practice endless number questions based on a given text.

Evaluation Service

A system designed to assess students' answers against stored answers, using NLP techniques, universally employed across all Embibe products for evaluation purposes.

RIO

Led the development of a product designed to automate digital advertising trading, using Multi-Armed Bandit algorithms to dynamically allocate media budgets across various strategies.

Predict

• A product tailored to predict prospect users to maximize the CVRs of ad campaigns using K-means and Random forest, processing terabytes of data. One of the best strategy chosen by all traders at MiQ to optimize CVR based campaigns.
• Tweaked existing model as feature selection model to select the best set of features for the campaign target optimization, post GDPR.

Cross Device

• A probabilistic model to generate the device graph which identifies Internet users existing across multiple channels for cross device tracking.
• Researched and evaluated cross device third party data to measure its impact in terms of insights and targeting.
• Automated a completed pipeline to generate various cross device based insights to understand campaign level cross device behaviors.

Digital Ad fraud detection

• Designed and Implemented an anomaly detection system which o proactively prevent digital Ad frauds.

Data Quality

• Developed a data quality solution at Impetus that performs data cleansing and deduplication.

Sentiment Analysis

• Developed a sentiment analyzer using data from telecom customer reviews.