Projects

Two Centuries of Sexism in British Parliament: A Computational Analysis of Women’s Representation in the Hansard Corpus

August 2025

Team: Mandira Sawkar, Omar Khursheed, Dr. Ashique KhudaBukhsh

Resources: [Code]

Summary: Built an LLM-as-judge pipeline to mine 6.7M UK parliamentary speeches for suffrage debates, achieving stance-classification agreement with human experts (k=0.71) that exceeded inter-annotator agreement itself (k=0.64). Designed a psychology-grounded sexism taxonomy and applied it at corpus scale, quantifying that anti-suffrage speeches were 2.6x more likely to contain sexist rhetoric (54% vs. 21%) and revealing a century-long shift from hostile to benevolent sexism..

My contribution: Designed the taxonomy and the human annotation protocol. Conducted prompt-tuning, statistical analysis of sexist rhetoric in the Hansard corpus and co-authored the resulting ARR paper.

Beyond Hallucinations: A Multi-Agent Approach To Fact-Checked AI Knowledge

May 2025

Team: Mandira Sawkar

Resources: [Code]

Summary: Developed a theoretical framework consisting of multiple LLM agents producing verifiable responses through debate and consensus.

My contribution: Sole contributor.

Detecting Deception: Intelligent Systems for Fighting Misinformation

May 2025

Team: Mandira Sawkar, Leona Joseph, Vivek Senthil

Resources: [Code]

Summary: Designed a logic-gated fact-checking system integrating RoBERTa classifiers with contradiction detection, validated on 7,200+ claims. Improved explainability and reliability by maintaining 83% accuracy while enabling selective overrides for low-confidence predictions.

My contribution: Conducted extensive literature review and developed a framework with a fine-tuned DistilRoBERTa with logic-gated overrides; Co-authored an ACM-style conference paper.

LLM-as-a-SupremeCourt-Judge

December 2024

Team: Mandira Sawkar

Resources: [Code]

Summary: Developed an AI framework with LLMs (GPT API, fine-tuning, chain-of-thought reasoning) to evaluate 200+ Supreme Court cases, revealing how factors such as panel size, prompting strategies, and ideology personas influence alignment between LLM reasoning and human judges.

My contribution: Sole contributor.

AI Alignment in Criminal Justice - Fair Jury Selection and Deliberation

December 2024

Team: Mandira Sawkar, Leona Joseph, Shrid Pant

Resources: [Code]

Summary: Created a theoretical hybrid model for fair jury selection and intra-jury deliberation using optimization and game theory with a novel extension to nash equilibrium.

My contribution: Conducted literature review and helped develop both the optimization and game theoretic components; Co-presented and won 1st place at the RIT Graduate Showcase.

Interpretable Parking Ticket Location Classifier

January 2023

Team: Mandira Sawkar

Resources: [Code]

Summary: Built a SparkML RandomForest classifier to predict the police-precinct of ticketed vehicles using the NYC parking-ticket dataset. Used PySpark for parallel computation. Achieved 98% accuracy, and identified police precincts with anomalous ticketing practices.

My contribution: Sole contributor.

Product Shipping Status Classifier

December 2022

Team: Mandira Sawkar

Resources: [Code]

Summary: Created a RandomForest model to predict issues in the product supplychain which might delay product shipment to customers. Employed PCA and LDA optimization for visualization and dimension reduction. RandomForest model achieved accuracy of 97% (with PCA) and 92% (with LDA).

My contribution: Sole contributor.

Cancer Detection as a Microservice

December 2022

Team: Mandira Sawkar

Resources: [Code]

Summary: Trained a Voting Classifier to predict cancer type with the Wisconsin Breast Cancer Dataset. Hosted the model as a Flask-based microservice running within a Docker container.

My contribution: Sole contributor.

getPaid

November 2022

Team: Mandira Sawkar

Resources: [Code]

Summary: Created a SparkML RandomForest model to predict total employee compensation. Queried data with SparkSQL, ran PySpark scripts to run EDA, pre-process data, and train model achieving with 0.98 R2 score.

My contribution: Sole contributor.