Daksha Ladia

Projects

• Research Study on Bias in Large Language Models

  Research Publication: Statistical Analysis, LLM Evaluation

  Conducted comprehensive analysis of pronoun and occupational biases across multiple LLMs using targeted prompts and statistical tests. Key findings: GPT-4o selected non-preferred pronouns <5% of the time, while Qwen2.5 exhibited 100% positional bias. Published detailed report with mitigation strategies, contributing to fair AI development.

• Research on Context Sensitivity in LLMs

  Research Focus: Large Language Models, Context Processing, Information Retrieval

  Investigated the "Stuck/Lost in the Middle" phenomenon in LLMs, analyzing how models handle contextual information across different positions in input sequences. This research provides insights into improving context comprehension and retrieval capabilities in modern language models.

• Multi-agent AI Browser Automation Platform

  Technologies: Vision LLMs, Playwright, Python, Multi-LLM Providers (Ollama, Gemini, OpenAI), Browser-user

  Built a vision LLM-powered browser automation system that navigates web applications (Notion, Asana) using screenshot analysis and DOM parsing to execute multi-step workflows autonomously. Architected a modular framework with multi-LLM provider support, hybrid state detection, and automated dataset generation for workflow capture and analysis. The system enables natural language-driven automation of complex web tasks without manual scripting.

• Recomm AI - Privacy-Preserving Shopping Recommendation Assistant

  Technologies: Differential Privacy, Federated Learning, PyTorch

  Developed a privacy-preserving recommendation system that protects user data while delivering personalized shopping recommendations. Evaluated the system against various privacy attacks, demonstrating robust protection without compromising recommendation quality—addressing the critical need for privacy in e-commerce AI systems.

• Secure, On-device, Privacy-first Personalized AI Companion

  Technologies: On-device ML, Federated Learning, Multimodal AI (Voice + Text)

  Developed a fully on-device AI companion that processes voice and text inputs locally, using federated learning for personalization without data leaving the device. This approach ensures complete privacy while delivering personalized experiences—addressing growing concerns about AI data privacy.

• Complex Query Synthesis for Enhanced Information Retrieval

  Technologies: LLM Fine-tuning, Document Segmentation, BM25, Python

  Developed an advanced document retrieval pipeline combining document segmentation, pseudo-query generation, and fine-tuned language models. Achieved 4% precision improvement over BM25 baseline, significantly enhancing search accuracy for complex queries in large document collections.

• Question-Answering Bot

  Technologies: Large Language Models, Document Processing, RAG

  Engineered an intelligent Q&A bot using advanced LLMs to provide precise, context-aware responses from document collections. Automates information retrieval processes, enabling users to quickly extract insights from large document repositories without manual searching.

• Real-Time Streaming & Data Prediction

  Technologies: Apache Kafka, Python, AWS, Docker, Neural Networks

  Implemented a real-time data streaming pipeline for trending GitHub repositories using Apache Kafka. Built neural network models to predict repository trends, enabling proactive discovery of emerging open-source projects. The system processes thousands of events per second with low latency.

• AI SlackBot

  Technologies: Slack API, LLMs, Python

  Built an intelligent Slackbot that automatically summarizes conversations and answers questions from Slack channel history. Enables team members to quickly catch up on discussions and find information without manually scrolling through thousands of messages, improving team productivity.

• Suno Na - Real-Time Language Translator

  Technologies: Real-time Translation, Speech Processing, Microsoft Garage Hackathon 2022

  Built a prototype real-time language translator designed for international conferences, enabling seamless communication during calls across different languages. The system addresses a critical need for global collaboration and accessibility in multilingual environments.

• Google Assistant Chatbot

  Technologies: DialogFlow, Google Assistant API, Multilingual NLP

  Created and deployed a multilingual chatbot (English and Hindi) on Google Assistant to help students discover and apply for travel grant scholarships for technology and business conferences. The chatbot has helped hundreds of students access funding opportunities they might have otherwise missed.

• Parallel Programming Research Implementation

  Technologies: CUDA, OpenMPI, C++

  Implemented a research paper algorithm for parallel odd-even merge sort using CUDA for GPU acceleration and OpenMPI for distributed computing. This project provided deep exposure to parallel programming paradigms and high-performance computing during undergraduate studies.

• Robowars 🏆 Gold Prize Winner

  Technologies: Robotics, Dijkstra's Algorithm, Path Planning

  Won the gold prize at Microsoft Garage Hackathon 2023 by developing an autonomous robot capable of exploring unknown environments, calculating optimal paths using Dijkstra's algorithm, and executing precise pick-and-drop operations. Demonstrated strong problem-solving and algorithmic thinking in a competitive setting.

• Health Care Portal

  Technologies: Full-stack Web Development, MFA, Payment Integration

  Built a comprehensive healthcare platform enabling doctors and patients to schedule appointments, process payments, and navigate using embedded maps. Implemented multi-factor authentication to secure sensitive medical data, creating a seamless and secure healthcare management experience.

Publications

• Virtual Community: An Open World for Humans, Robots, and Society

  Technologies: Unity, Python, Multi-Agent Systems, Embodied AI

  Developed a comprehensive simulation platform for multi-agent embodied AI using Unity and Python. Generated benchmarks for speed and resolution across different robots, agents, and scenes with varying parameters. This platform enables researchers to test and evaluate AI agents in realistic virtual environments, advancing research in embodied AI and human-robot interaction.

  Status: Under review at ICLR 2026

• Do Different Models Differ in their Bias Towards Occupational Preferences?

  Research Focus: LLM Bias Analysis, Statistical Evaluation, Fair AI

  Conducted comprehensive analysis of pronoun and occupational biases across multiple Large Language Models using targeted prompts and rigorous statistical tests. Key findings: GPT-4o demonstrated strong performance with non-preferred pronoun selection <5% of the time, while Qwen2.5 exhibited 100% positional bias. Published detailed report with actionable mitigation strategies to improve fairness in LLM applications.

  Status: Submitted to COLM'25

Research Interests

My research focuses on advancing Context Engineering, Post Training, LLM Inference, Multi-Modal LLMs and Multi-Agent Autonomous Systems, with particular interest in:

• • Multi-agent coordination and communication in complex environments
• • Bias detection and mitigation in large language models
• • Privacy-preserving machine learning systems
• • Context sensitivity and information retrieval in LLMs

Work Experience

• Machine Learning Engineer Intern, System1 | May 2025 - Aug 2025

  Building a multi-agent AI pipeline for SQL generation from Natural Language queries with an evaluation framework, achieving a 98% success rate. The system enables non-technical users to query databases using natural language, significantly reducing the barrier to data access.

• Intelligent Healthcare Document Processing Platform | Freelance Project | June 2024 - August 2024

  Technologies: GPT-4o, Google Vision API, Python, OCR

  Built and successfully sold enterprise software to a third-party medical insurance company. Architected an end-to-end AI-powered pipeline that automates healthcare document classification, OCR processing, and conversion of unstructured data into structured formats. The solution reduced manual processing time by 99% and enabled the client to process thousands of documents daily with high accuracy, transforming their operational efficiency and reducing costs.

• Software Engineer, Microsoft | June 2020 - July 2024

  Bing Ads Team — Optimized ad ranking systems serving millions of users globally:

  • Improved ad ranking runtime and accuracy while reducing infrastructure costs through optimized selective sampling
  • Built automated Regression Tester Tool, saving 90% of developer validation time
  • Enhanced ML models for ad performance prediction, achieving 2-4% AUC improvements and driving measurable revenue growth
  • Executed strategic optimizations across international markets, improving ad relevance and user engagement

• Data Science Intern, FarmGuide (Acquired by DeHaat) | Jan 2020 - May 2020

  Developed ML models for crop identification and health monitoring with 96% accuracy. Built forecasting algorithms using ARIMA and Prophet to provide environmental insights and crop recommendations, helping farmers maximize revenue through data-driven decisions.

• Software Engineering Intern, Microsoft | May 2019 - July 2019

  Analyzed seasonality patterns and anomalies in Bing advertiser campaigns using time-series modeling. Optimized bid pricing algorithms, improving ad performance metrics during peak holiday periods and increasing campaign effectiveness.

• Mentor, CodeYoung | 2019 - 2020

  Taught Python and web development to students globally, focusing on practical applications. Mentored students in building real-world projects, with one student securing admission to Yale Summer School based on my recommendation.

• Software Developer Intern, Cygnet Infotech | 2018

  Developed a secure employee onboarding application using .NET and AES encryption, ensuring complete confidentiality for sensitive personal data and documents in enterprise environments.

Honors, Awards & Others

• Innovation & Excellence Award, Microsoft

  Received in FY20–FY21 for leading a pipeline migration project that streamlined workflows and significantly improved operational efficiency.

• Judge, MITHacks 2025

  Judged at MITs hackathon for undergraduate students, in 2025.

• Reviewer, GHC 2025

  Helped with the selection of applications submitted for different verticals

• Mentor, HackHarvard

  Mentored undergraduate students at Harvard's undergraduate hackathon in 2024, supporting innovation and technical growth.

• Participant, MIT Policy Hackathon

  Participated and proposed an interpretable solution for analyzing error rates in SpO2 measurements during the competition, in 2024.

• Mentor, She Hacks DTU

  Mentored participants at Delhi's largest all-women hackathon in 2021, fostering innovation and collaboration among attendees.

• Top 1% in Higher Secondary Exams

  Achieved a top percentile rank in national exams in India and received the Inspire Scholarship from the Indian government in 2015 for academic excellence.

Contact

I'm always open to discussing research opportunities, startups, collaborations, or interesting projects. Feel free to reach out!

dladia@umass.edu

Gallery

Glimpses of some moments...