Under Submission · COLM 2026
CircuitSteer: Circuit Discovery-Based Steering via Sparse Autoencoder Feature Flow
Available upon request
01
About
I am a machine learning researcher and engineer completing an M.S. in Computer Science at the University of Southern California, with a background in Computer Engineering from Shahid Beheshti University.
My work spans deep learning theory, bioinformatics, and medical AI, with a focus on multiview and multimodal learning, generative and probabilistic models, and graph neural networks — leading to peer-reviewed publications in Nature Communications (2025) and IEEE Access (2024).
More recently I have concentrated on large language models, particularly alignment, safety, and mechanistic interpretability. My ongoing work explores circuit-level understanding and behavioral steering of LLMs, including multi-layer steering methods inspired by transformer circuit discovery.
In parallel, I build end-to-end ML systems across multimodal RAG, LLM evaluation, and deep research agent pipelines. I am particularly interested in bridging theory and real-world systems — taking insights from representation learning and interpretability into practical, scalable AI.
Education
University of Southern California
MS Computer Science · Jan 2024 – Dec 2025
Shahid Beheshti University
BSc Computer Engineering · Sep 2018 – Feb 2023
Technologies
02
Research & Publications
Nature Communications · 2025
Multi-organ metabolome biological age & cardiometabolic risk
Under Submission
EvoGLAD: Evolving Graph-Level Representations for Anomaly Detection in Metagenomic Communities
Available upon request
arXiv Preprint · 2025
Enhancing Multi-Modal Video Sentiment Classification Through Semi-Supervised Clustering
03
Experience
Jan 2025 – Present
Research Intern
Dr. Neiswanger's Lab · USC
Graph representations for metagenomic communities (METAGENE-1). Developing EvoGLAD, a framework for graph-level anomaly detection in microbial ecology using evolutionary graph neural networks.
May 2024 – Present
Research Intern
LABS · Columbia University & USC
Disease subtype discovery via weakly supervised deep clustering. Collaborating on multi-organ metabolomics research published in Nature Communications, applying multiview learning to large-scale biobank data.
Apr 2024 – Dec 2025
AI Engineer
USC CARC
Developing a novel LLM-based approach for deep academic research. Building end-to-end pipelines combining multimodal RAG, LLM evaluation and benchmarking, and agentic workflows for scientific discovery.