Aditi's Portfolio

Publications

Spivak, Shani C., Luca Podo, Rohit Gandikota, Aditi Krishna, Enrico Bertini, and Melanie Tory. 2024. “Characterizing LLM Visualization Errors.” OSF. August 23. doi:10.17605/OSF.IO/WZQ3G.

Developed an LLM-based chat interface using Streamlit, GPT-3.5, and Microsoft Presidio API, incorporating behavioral economics principles and visual nudging elements to improve user privacy awareness in AI interactions
Conducted pilot user study with semi-structured interviews and think-aloud protocols, finding that visual nudging elements enhanced privacy awareness and data disclosure behavior in 67% of participants
Performed thematic analysis identifying 5 key themes, and 16 sub-codes around user privacy strategies, revealing that participants employ selective sharing, fake personas, and trust-based decisions when interacting with AI systems

Developed a hybrid model combining 3D-CNN and LSTM networks that fused meteorological data and radar images to predict snow over Lake Michigan with an F1 score of 0.87 - a 117% improvement over the baseline decision tree model
Created a novel data preparation function to transform the numerical data into sliding window of observations to predict snow 3 days in advance based on 5 days of historical data

Engineered and deployed T5 and BERT language models on AWS Lambda for real-time summarization and sentiment analysis of web-scraped stock news, orchestrating daily update pipelines using Apache Airflow
Designed an intuitive user interface for collated financial news, integrating color-encoded sentiment indicators to enhance information comprehension
Enabled user logging in BigQuery to analyse user engagement statistics through an interactive dashboard on Data Studio
Conducted preliminary user interviews with 5 graduate students, demonstrating the effectiveness of the design intervention as 80% of participants reported it significantly influenced their stock trading decisions

Utilized unsupervised learning techniques including k-means, hierarchical clustering, and DBSCAN to segment customers into four key demographic groups
Leveraged principal component analysis to reduce feature dimensions and boost model performance
Assessed models via silhouette analysis, determining agglomerative hierarchical clustering as the top performer with a silhouette score of 0.42

Designed and implemented a Python app with Streamlit using SEVIR (Storm EVent ImageRy) dataset to enable a deep-learning-based Nowcasting API built with FastAPI web framework
Built an API that generated and cached the GIF of the near-term forecasted storm images to Google Cloud Storage
Deployed large-scale deep learning HuggingFace T5 model on service endpoint using AWS Lambda functions to summarize the lat-long specific storm event narrative from the NOAA dataset
Enabled batch requests for generating images for multiple locations for hourly updates using Apache Airflow, and cached the output to Google Cloud Storage
Enabled user logging in BigQuery to analyse user engagement statistics through an interactive dashboard on Google DataStudio, and hosted the application on Google App Engine

Implemented Similarity Search Algorithm built by Ilya Katsov
Enabled data ingestion and search on Elastic Cloud using Python Client and Kibana Console
Build a Streamlit web interface to allow users to choose the product of their liking and view k-similar images that they might like

Designed and implemented a relational database using SQL to build a unified system for hotel businesses to enable reservation management of rooms, restaurants, and events
Created triggers to compute membership status based on total points earned by customer, and final amount due on invoice for each reservation
Built a dashboard in Tableau to visualize the performance of various revenue segments, such as memberships and events

Built a web app using Streamlit to provide product recommendations based on association rules generated using Apriori Algorithm for similar customers
Performed Market Basket Analysis to develop more effective product placement, pricing, cross-sell, and up-sell strategies on retail dataset
Identified target customer segments using RFM analysis to build promotional strategies
Provide an interactive dashboard to analyze country-level performance of products, customers, and sales revenue statistics on a monthly, weekly, and daily level