Insights Generation on Equifax Time Series Data using Generative AI

Developed an automated system using Generative AI handling 1M+ time-series data points, enabling real-time detection of key financial deviations, thereby improving the accuracy and speed of Equifax’s customer credit assessments. Implemented Change Detection algorithms to monitor credit metrics.

Purdue Data4Good Case Competition - Question/Answering system using LLMs

Implemented a Question/Answering system with open-source LLM models (GPT, Llama, MPT). Utilized zero-shot prompt engineering on doctor-patient dialogues and dictations for precise information extraction. We also integrated Prediction Guard API and Microsoft Azure ML into the project workflow.

Recipe Recommendation using Multi-modal embeddings

Enhanced a state-of-the-art recipe recommendation system by leveraging embeddings on text data as seen in traditional recommender systems and Residual Network (ResNet) embeddings for visual information. Incorporated these embeddings into a graph neural network-based model called Recipe2Vec

Analyzing intrinsic features of articles using Twitter popularity

Engineered a sophisticated visualization tool to unveil topic similarities based on keywords and investigate the intricate relationships between article popularity and its intrinsic features such as metadata, sentiment, TF-IDF score, abstract length, print headlines sentiment, and word count.

Forecasting Weekly Sales of Low-Selling SKUs

Create a forecast for slow-selling SKUs for one week out that champions both accuracy and speed. Employed prompt engineering techniques and experimented with various time series and ensemble methods to achieve a low RMSE of 3.30 using a Random Forest model trained in just 32 seconds.

Core Temperature Estimation of EV Battery Packs

Core temperature knowledge is critical for ensuring safety. In smaller (gadget) batteries, the easily measured case temperature is strongly correlated with core temperature. However, this is not true of larger EV batteries. The project aims to predict the core temperature of an electric vehicle.

Network Monitoring Optimization

Recent incidents underscore infrastructure vulnerabilities. In this project, a water utility company, and more specifically a network operator, is interested in allocating pressure sensors on nodes to detect pipe bursts in its water distribution. The goal is to maximize pipe coverage efficiently.