Automatically Detecting Numerical Instability in Machine Learning Applications via Soft Assertions
FSE 2025 paper introducing Soft Assertions to detect and trigger numerical instability bugs in ML applications.
Anwar Zahid
Ph.D. student working on reliable machine learning and software systems.
I study how AI and numerical software fail in practice, then build testing and debugging tools that make those failures easier to detect, reproduce, and fix.
About
I am a Ph.D. student in Computer Science at Iowa State University, advised by Prof. Wei Le in the Program Analysis and AI Lab. My research focuses on reliable AI systems, numerical debugging, and software engineering techniques for machine learning systems.
Before starting my Ph.D., I worked as a software engineer on government, banking, AI, and mobile platforms. That industry background shapes how I approach research: I care about methods that can become practical tools for developers and researchers.
Research Interests
- LLM hallucination
- AI reliability
- Software engineering for AI systems
- Interpretable / white-box analysis
- Human-centered AI tools
Projects
Selected projects from existing repository entries.
GPU Numerical Testing – HIP Code Generation in Varity
Contributed to the LLNL Varity project by implementing HIP backend generation for GPU kernel testing, enabling cross-platform numerical consistency evaluation.
Hate Speech Detection with Geographical Context
Extended a class project to evaluate hate speech detection models using geographical metadata. Later adapted for large language model testing and published in 2025.
Face Recognition with Liveliness Detection
Developed a real-time face recognition and spoof detection system for FinTech applications, enhancing security for remote banking verification.
Parliament Resource Planning – SSO Integration
Implemented Single Sign-On authentication for PRP system, enabling secure unified access across parliamentary resource modules.
COM S 572 – Machine Learning Project
Implemented and evaluated ML algorithms for natural language processing tasks. Later extended for LLM experiments and paper publication.
Othello AI Agent
Developed an intelligent Othello game-playing agent using adversarial search algorithms and heuristic evaluation functions.
Mancala Game Simulation
Created a Mancala game engine with an AI opponent using minimax and heuristic strategies.
Ray Tracing Renderer
Built a ray tracing engine from scratch to render 3D scenes with reflection, refraction, and shading.
Nachos – Instructional Operating System
Extended the Nachos instructional OS to implement thread scheduling, virtual memory, and file system operations.
Channel Equalization for Wireless Communication
Implemented and tested channel equalization algorithms to improve signal quality in noisy communication environments.
Publications
Peer-reviewed and preprint work on ML reliability, LLM evaluation, and numerical correctness.
Automatically Detecting Numerical Instability in Machine Learning Applications via Soft Assertions
Introduces Soft Assertions, a method for detecting and triggering hidden numerical instability bugs in machine learning applications.
Evaluation of Hate Speech Detection Using Large Language Models and Geographical Contextualization
Evaluates how large language models perform on hate speech detection when geographic and social context are included.
Testing GPU Numerics: Finding Numerical Differences Between NVIDIA and AMD GPUs
Studies numerical differences between NVIDIA and AMD GPU executions and their implications for reproducibility and portability.
A Conceptual Design of Virtual Internship System
Presents a conceptual software platform for virtual internship delivery and software development skill benchmarking.
Writing
Short research notes and engineering write-ups.
What I Want This Blog to Become
I want this blog to be a working notebook for the problems I keep returning to: machine learning reliability, numerical instability, debugging, and the engineering dec...
Automatically Detecting Numerical Instability in ML via Soft Assertions
Machine learning (ML) models run on massive datasets and often perform billions of floating-point calculations. But here’s the problem: small numerical errors can snow...
Testing GPU Numerics: Finding Numerical Differences Between NVIDIA and AMD GPUs
When you run the same GPU program on an NVIDIA GPU and an AMD GPU, you might expect identical results. Surprisingly, that’s not always the case — even small floating-p...
Contact
The easiest way to reach me is by email. You can also find my code and academic profiles through the links below.