aryan singhal
artificial intelligence at Purdue. building Mirrors.
i study artificial intelligence at Purdue. right now i'm building Mirrors, a staging environment for AI agents. it rebuilds the systems your agents call, then replays real sessions against them, so you catch regressions before your users do.
before that i spent a few years on ML research: temporal graph networks at UT Austin, hate speech severity modeling at UC Santa Barbara, and a quantum vision transformer that ended up at NeurIPS 2024.
work
Mirrors 2026 to present
a staging environment for AI agents. most teams have no test instance for the internal systems their agents depend on, so Mirrors learns those systems from traces, code, or docs, stands up a runnable replica, and replays real sessions against it to catch regressions before they ship.
Profine 2026
automatic GPU profiling and optimization for PyTorch training. it profiled training code on real hardware, applied validated optimizations, and benchmarked what actually got faster. not working on it anymore.
projects
QViSTA 2023 to 2025
a hybrid quantum-classical vision transformer for multi-stage Alzheimer's classification from MRI. it swaps parts of a standard ViT for variational quantum circuits, cutting parameter count while holding accuracy. published at NeurIPS 2024.
temporal GNN for political polarization 2024 to 2025
with researchers at UT Austin. a multi-modal temporal graph neural network over 2K+ subreddits that models how users move between communities and how their topics shift, so polarization shows up as a trajectory instead of a snapshot.
hate speech ensemble model 2024 to 2025
with researchers at UC Santa Barbara. an ensemble that scores hate speech by severity on a 1 to 6 scale instead of flattening it into one label, combining tree-based and neural models with a semi-supervised pipeline for scarce, imbalanced data.
multilingual claim verification 2025
a fact-checking pipeline that evaluates LLMs across 15 languages using direct inference, self-translation, and pre-translation, measuring exactly where models degrade by language family and model size.
Forkprint
a web app for tracking food expiration and logging waste, with suggestions for using what you already have. built to make household food waste concrete enough to actually change.
papers
- QViSTA: A Novel Quantum Vision Transformer for Early Multi-Stage Alzheimer's Diagnosis Using Optimized Variational Quantum Circuits
- Can LLMs Verify Arabic Claims? Evaluating the Arabic Fact-Checking Abilities of Multilingual LLMs
- A Comparative Study of Translation Bias and Accuracy in Multilingual Large Language Models for Cross-Language Claim Verification
- Multilingual Fact-Checking using LLMs
- Unveiling Bias in ChatGPT-3.5: Analyzing Constitutional AI Principles for Politically Biased Responses
- Mitigating Political Bias in Large Language Models Using Chain of Thought Prompting Techniques
- Inferring Hate Speech Trends for Contemporary Tweets Using a Novel Machine Learning Approach from Supervised Learning Algorithms
background
elsewhere
github linkedin x scholar letterboxd
best way to reach me is aryan.singhal2016@gmail.com.
