AimBot: An investigation into external memory and advanced prompting for LLM agents

Mehul Maheshwari, Sai Chaparala

Finalist in AWS X VCT Hackathon.

Dec 2024

Paper / Demo / Devpost

Large Language Models (LLMs) have trended towards increasing parameter sizes to improve their performance across a wide range of tasks. However, recent advances demonstrate that smallerscale models, when combined with knowledge bases (KBs) and effective prompting strategies, can achieve comparable results while being more efficient and interpretable (Borgeaud et al., 2022). This project investigates strategies for integrating KBs and optimized prompting techniques in generating high-quality, task-specific responses.

Can Item Metadata Combat the Cold Start Problem?

Mehul Maheshwari, Jerry Thomas, Adrian Ong, You Wu

Dec 2024

Proposed new recommender systems beating existing baselines by taking advantage of metadata on the Steam User-Game dataset.

Abstract: The cold start problem in recommender systems arises when dealing with new items with no prior interaction history, making it challenging to generate accurate recommendations. This work ex- plores whether incorporating item metadata can effectively address this problem in game recommendations using user-game interactions on the Steam platform. We leverage game metadata (genres and price) alongside implicit feedback from user playtime data, comparing five different recommendation approaches: traditional Matrix Factorization (SVD), Factorization Machines (FM), hybrid methods combining SVD/FM with K-Nearest Neighbors (KNN), and a metadata-only approach using FM. An analysis of results show that the SVD baseline shows strong performance on existing games but degrades for new games, demonstrating the cold start problem. Surprisingly, incorporating metadata through FMs doesn’t improve performance, suggesting that with the Steam dataset’s scale (>5 million interactions), the collaborative filtering signal alone may be sufficiently strong. The hybrid method which incorporate game metadata (Method III) show significant improvements on items prone to cold start problems through extending their base counterparts, while the metadata-only approach (Method V) performs similarly to Method II and IV, indicating that in this context, item metadata provides limited additional signal beyond user-game in- teractions.

Discerning Causes of Ratings Bias: A Platform for Bias Experimentation in Ratings-Based Reputation Systems

Published in the 12th Association for the Advancement of Artificial Intelligence Conference on Human Computation and Crowdsourcing, Aug. 2024

August 2023

As the gig and sharing economy proliferates, issues of demographic disadvantages in ratings-based reputation systems become critically important. After all, if women or non-white workers receive unfairly biased ratings, they will be disadvantaged in these kinds of marketplaces. Some prior observational work identifies these biases in the real world, whereas other work does not replicate these biases experimentally. To help characterize how and why these biases play out, we introduce a platform for streamlining the process of creating experiments in the space of five-star ratings. We aim to use this platform as a stepping stone to investigate which analytical dimensions cause and mitigate ratings bias.

Tech Stack: Django, Python, Next.js, AWS Tools: EC2 + Route53, SQLite, PostgreSQL, Gunicorn, Nginx, PM2

Sign Langauge Learner

We created an AI model that can translate sign language into text. Although models already exist, we experimented with various combination of data augmentation, transfer learning, and model architecture. Our final product greatly surpasses existing baseline accuracies.

Tech Stack: Tensorflow, PyTorch, Python, Juypter, Keras, Kaggle

Madtown Bhangra Website

This is a website I developed for the Madtown Bhangra competition to help garner attention from sponsors and centralize information for prospective teams and audience members.

Tech Stack: HTML, CSS, Javascript, Boostrap, GitPages

Personal Website

I've always wanted my own website, a little digital testament to myself. One day I decided to turn that dream into a reality.

Tech Stack: HTML5, CSS, Javascript, GitPages

Cryptography with Chaotic Systems

I created an encryption/decryption algorithm utilizing chaotic ordinary differential equations, which are extremely difficult to crack due to their turbulent nature. Additionaly I wrote a report on the nature of these encryption algorithms, as well as additional steps to bolster their security.

Tech Stack / Skills: Python, Chaotic Differential Equations, Research, Analysis of Differential Systems

Badminton Club @ UW-Madison Website

As the Media and Communications Chair of the badminton club at UW-Madison, I noticed that the club struggled to communicate information that all club members wanted access to, such as pictures or upcoming practice schedules. I decided to tackle this problem by leveraging my CS background, and created a website to communicate this information more readily.

Tech Stack: HTML5, CSS, Javascript, GitPages

Office Hours Waitlist

We created this application to help manage office hours queues, and cut down on waiting times. I worked on this full stack application alongside 3 of my peers.

Although a minimum viable product exists, unfortunately this project did not complete. However, it was a useful learning opportunity for understanding full stack development.

Tech Stack / Skills: React, Python, MySQL Workbench, Full-Stack Development

Computer Graphics Investigation

These are a series of projects I worked on to pursue the Graphics sect of CS. I worked methodologically, with each project increasing in complexity and tackling a new dimension (both metaphorically, and in project 4 onwards literally) of graphics.

Topics Covered: 3D -> 2D Projected Views, Hierarchical Modeling, Parametrics in Animation, Transformations using Explicit Matrix and Vector Representations

Sorting Visualizer

I wanted to get some practice using React, as well as refreshing my memory on important algorithms and tackled both problems via this project.

Tech Stack / Skills: React, Sorting Algorithms