Anna Mazhar
Anna Mazhar
Ph.D. Candidate
Cornell University

Hi! I am Anna, a second-year PhD student at Cornell University, advised by Sainyam Galhotra. My research lies at the intersection of systems, software engineering, and machine learning, focusing on the rigorous evaluation of complex systems with AI-driven components. I aim to develop principled techniques to improve system correctness, reliability, and trustworthiness in practice.

Before starting at Cornell, I completed my MS in Computer Science at UIUC, advised by Tianyin Xu. There, my research centered around cloud systems reliability.

Education

  • Cornell University
    Ph.D. in Computer Science
    2024 – Present
  • UIUC
    MS in Computer Science
    2022 – 2024
  • LUMS
    BS in Computer Science
    2018 – 2022

Research

Published Work

Trace-Level Analysis of Information Contamination in Multi-Agent Systems
Anna Mazhar, Huzaifa Suri, and Sainyam Galhotra
ACM Conference on AI and Agentic Systems (CAIS'26)
pagepdf
Towards Reliable Testing of Machine Unlearning
Anna Mazhar, and Sainyam Galhotra
ACM International Conference on the Foundations of Software Engineering (FSE'26 – IVR)
pdf
Fidelity of Cloud Emulators: The Imitation Game of Testing Cloud-based Software
Anna Mazhar, Saad Sher Alam, Xinze Zheng, Yinfang Chen, Suman Nath, and Tianyin Xu
IEEE/ACM International Conference on Software Engineering (ICSE'25)
pdfbibslidestalk

Manuscripts in Preparation

Causal Fuzzing for Verifying Machine Unlearning
Anna Mazhar, and Sainyam Galhotra
Preprint [arXiv:2509.16525]
pdfarXiv

Selected Projects

  • Evidence Degradation in Multi-Agent Workflows
    Analyzing evidence degradation in multi-agent workflows using perturbation-based evaluation. Demonstrated that minor corruptions in inputs (e.g., OCR errors, table misalignment) systematically alter workflow traces, with implications for reliability in retrieval-based AI systems.
  • Causal Verification Framework for Machine Unlearning
    Designed a causal influence-based verification framework for Machine Unlearning that detects hidden data residuals missed by existing approaches, advancing the debuggability of ML systems.
  • Cloud Emulator Fidelity Testing
    Led the first systematic study of behavioral fidelity in cloud emulators (Azurite, LocalStack), uncovering discrepancies in 94 of 255 APIs across AWS and Azure via a custom SDK fuzzer. Resulted in 12 confirmed bugs and 6 upstream fixes.

Selected Awards

  • SIGSOFT Travel Grant — Travel support for attending conferences
  • Erasmus Mundus Scholarship (€49,000) — Awarded to 26 applicants out of 735
  • Summer@EPFL Fellowship — Selected from ~4,500 applicants with a 1–2% acceptance rate