AI assistants are increasingly prevalent in the completion of everyday tasks such as generating sentences and sending emails. However, they suffer from limitations such as poor user personalization, outputs subject to developer biases, and security risks during data transfer. The current Panza system can generate emails in a user’s writing style based on a given prompt using data playback and a RAG with RoSA to train model parameters. I will work on Panza’s next portion, a Panza associative memory module; this work applies the principles of personalization and privacy to create a smart search locally trained on all of a user’s personal data across various platforms (such as emails, Slack/text messages, search history, documents, etc.) to provide a complete view of a user’s device usage. The novelty of this method lies in the connections between the data: instead of a simple RAG database that stores information which is retrieved based on the embeddings’ similarity to a query, we will create an associative memory module that relates relevant data based on the time they were accessed. In this way, we can create a search tool that performs well even when only given a vague recollection of details, allowing users to save hours of time combing through old records and files instead of simply asking a system.

The SuperUROP program will allow me to deepen my knowledge in machine learning and take ownership of a project. Having built a machine learning tool in an internship and taken related classes, I hope to apply my skills to advanced research, further my understanding of developing intelligent systems, and provide meaningful contributions to the field. The hands-on experience gained will prepare me to tackle future engineering challenges.