About me

I'm a computer science master's student at the University of North Carolina at Chapel Hill, where I work with Dr. Leonard McMillan, whose research spans computer vision, image processing, and computational genomics. My own work sits in computer vision and efficient machine learning: getting capable models to run well on the hardware people already own.

Before Chapel Hill, I did my undergrad at UNC Charlotte, where I got hooked on research early. I analyzed job traces from large-scale computing clusters in the High-Performance Computing Lab with Dr. Dong Dai, and built efficient video understanding models in the Machine Learning Lab with Dr. Srijan Das. That work grew into publications at CVPR 2026, npj Computational Materials, JSSPP'24, and IPDPS'24.

Outside the lab, I teach as a TA for DATA 120: Ethics of Data Science and AI, and I'm usually building something on the side. Lately that is Swiftlet, an open-source Swift and Metal runtime that streams 35B and 80B mixture-of-experts models on ordinary Apple devices, including iPhones. What drives all of it is simple: I like taking ideas out of papers and turning them into things people can actually use.

Peer Reviewed Selected Publications

  • MS-Temba: Multi-Scale Temporal Mamba for Efficient Temporal Action Detection
    Arkaprava Sinha, Monish Soundar Raj, Pu Wang, Ahmed Helmy, Srijan Das
    Published at CVPR 2026.

    View Paper Project / GitHub

  • AI-Assisted Rapid Crystal Structure Generation Towards a Target Local Environment
    Osman Goni Ridwan, Sylvain Pitié, Monish Soundar Raj, Dong Dai, Gilles Frapper, Hongfei Xue, Qiang Zhu
    Published at npj Computational Materials, 2026.

    View Paper

  • An Empirical Study of Machine Learning-based Synthetic Job Trace Generation Methods.
    Monish Soundar Raj, Thomas MacDougall, Di Zhang, and Dong Dai
    Published at 27th JSSPP’24 (Job Scheduling Strategies for Parallel Processing), 2024.

    View Paper GitHub Repository

  • Cross-System Analysis of Job Characterization and Scheduling in Large-Scale Computing Clusters.
    Di Zhang, Monish Soundar Raj, Bing Xie, Sheng Di, Dong Dai
    Published at 38th IPDPS'24 (IEEE International Parallel and Distributed Processing Symposium), 2024.
    Joint work by researchers from UNC Charlotte, Microsoft, and Argonne National Laboratory.

    View Paper Web App GitHub Repository

News

  • [2026] Started working with Dr. Leonard McMillan at UNC Chapel Hill, whose research spans computer vision, imaging, and computational genomics.

  • [Aug. 2026] Released Swiftlet, an open-source Swift and Metal runtime that streams 35B and 80B Qwen MoE models on Apple devices, including iPhone. Now past 650 stars on GitHub.

  • [Feb. 2026] 1 Paper - MS-Temba: Multi-Scale Temporal Mamba for Efficient Temporal Action Detection accepted to CVPR 2026, Website.

  • [Aug. 2025] Joined the Society-Centered AI Lab (SAIL) at UNC Chapel Hill, working with Dr. Neil Gaikwad on Human-AI Value Alignment and Computer Vision research.

  • [Jun. 2025] 1 Poster - Crystal Structure Prediction with Deep Learning Models Using Space Group Symmetry presented at Gordon Research Conference.

  • [Jan. 2025] Started working as a Teaching Assistant for DATA 120: Ethics of Data Science and AI at UNC Chapel Hill.

  • [Jan. 2025] 1 Paper - MS-Temba: Multi-Scale Temporal Mamba for Efficient Temporal Action Detection is out, Check Here.

  • [Dec. 2024] 🎉 Excited to share that I made the Chancellor's List for Fall 2024!

Research Posters

  • Crystal Structure Prediction with Deep Learning Models Using Space Group Symmetry
    Osman Goni Ridwan, Monish Soundar Raj, Md Shah Nawaz Arafin, Dong Dai, Qiang Zhu
    Presented at Research at High Pressure Gordon Research Conference.
    📅 This work will be publicly released after the main publication is finalized.

  • Enhancing HPC Job Scheduling with Synthetic Data Generation for RL-Based Schedulers
    Monish Soundar Raj, Dong Dai
    Presented at Undergraduate Research Summer Research Symposium.

    View Poster

Mentoring Activities

Awards

  • TCPP Travel Grant Award $700
    To attend and present at IPDPS'24, San Francisco, CA..

  • Office of Undergrad Research Travel Grant Award $500
    For travel expenses to attend IPDPS'24, San Francisco, CA

Charlotte

Resume

Experience

  1. Graduate Research Assistant - UNC Chapel Hill Computer Science

    2026 - Present
    • Working with Dr. Leonard McMillan, whose research spans computer vision, image processing, data-driven modeling, and computational genomics.

  2. Graduate Research Assistant - Society-Centered AI Lab (SAIL)

    Aug 2025 - 2026
    • Worked with Dr. Neil Gaikwad on Human-AI Value Alignment and Computer Vision research.

  3. Graduate Teaching Assistant - DATA 120: Ethics of AI and Societal Decision Making

    Aug 2025 - Present
    • TA for DATA 120, working with Dr. Hsun-Ta Hsu and Dr. Justin Sola, supporting 100+ students exploring the ethical dimensions of AI systems and data-driven decision making.
    • Design and lead interactive workshops on AI ethics, facilitating discussions on algorithmic fairness, bias mitigation, and responsible AI deployment in high-stakes domains.
    • Grade assignments and provide substantive feedback on student analyses of ethical case studies involving AI in healthcare, criminal justice, and social systems.

  4. Research Assistant - Charlotte Machine Learning Lab

    Aug 2024 - May 2025
    • Worked on computer vision and robotics research under Dr. Srijan Das.
    • Vision-Language-Action robotics: deployed a UR5e robotic arm with ROS 2, Real-Time Data Exchange integration, and a camera calibration pipeline for real-time visual perception, then implemented a vision-to-action pipeline using the LLaRA model for natural language-driven manipulation.
    • MS-Temba (second author, CVPR 2026): developed the CLIP feature extraction pipeline processing 12,000+ videos and 16M+ frames, and optimized baseline models for a 24% accuracy improvement on the Toyota Smarthome dataset.

  5. Research Assistant - High Performance Computing and Systems Lab

    Aug 2023 - May 2025
    • JSSPP'24 (first author): led "An Empirical Study of Machine Learning-based Synthetic Job Trace Generation Methods", integrating five generative models (GAN, CTGAN, TVAE, Gaussian Copula, Copula GAN) and demonstrating the superiority of ML models over statistical methods for high-fidelity trace generation.
    • IPDPS'24 (second author): contributed to "Cross-System Analysis of Job Characterization and Scheduling in Large-Scale Computing Clusters", analyzing job traces from the Theta, Supercloud, and ThetaGPU clusters for scheduling optimization.

  6. Web Developer - High Performance Computing and Systems Lab

    Jun 2023 - Aug 2023
    • Developed a web application visualizing 20+ job trace cluster characteristics, processing 1 million+ job log rows in real time.
    • Built with Python, Streamlit, Pandas, and NumPy.
    View Project: Lumos-Job-Traces

Education

  1. Master of Science in Computer Science

    Currently pursuing my master's degree at the University of North Carolina at Chapel Hill. My coursework and research focus on computer vision, machine learning, and efficient on-device inference. Working with Dr. Leonard McMillan.

  2. Bachelor of Science in Computer Science

    Completed my bachelor's degree at the University of North Carolina - Charlotte with Chancellor's List honors. My coursework and research focused on leveraging advanced programming tools, artificial intelligence, high-performance computing, and machine learning for data analysis and computational model development.

Technologies & Tools

Python Python
TypeScript TypeScript
Swift Swift
JavaScript JavaScript
HTML HTML
CSS CSS
Linux Linux
Windows Windows
React React
Node.js Node.js
Express.js Express.js
Next.js Next.js
Python Python
TypeScript TypeScript
Swift Swift
JavaScript JavaScript
HTML HTML
CSS CSS
Linux Linux
Windows Windows
React React
Node.js Node.js
Express.js Express.js
Next.js Next.js
Flask Flask
Streamlit Streamlit
MongoDB MongoDB
Firebase Firebase
NumPy NumPy
pandas pandas
Matplotlib Matplotlib
Heroku Heroku
AWS AWS
Flask Flask
Streamlit Streamlit
MongoDB MongoDB
Firebase Firebase
NumPy NumPy
pandas pandas
Matplotlib Matplotlib
Heroku Heroku
AWS AWS
View Full CV (Being Cooked)

Research

  • MS-Temba figure: a 40 minute activities of daily living video with short and long overlapping actions, and MS-Temba beating transformer models with far fewer parameters

    CVPR 2026 Conference paper

    MS-Temba: Multi-Scale Temporal Mamba for Understanding Long Untrimmed Videos

    Arkaprava Sinha, Monish Soundar Raj, Pu Wang, Ahmed Helmy, Hieu Le, Srijan Das

    17 citations 2026

    Temporal action detection in long, untrimmed videos of daily living, where a 40 minute recording mixes short actions like drinking from a cup with long ones like using a laptop, and many of them overlap. MS-Temba extends Mamba with dilated temporal state space blocks that read the video at several time scales and a scale-aware fuser that merges them. It reaches state of the art on the ADL benchmarks with about 5x fewer parameters than transformer models.

    Mamba State Space Models Temporal Action Detection Video Understanding PyTorch
    Paper Code
  • LEGO-xtal figure: symmetry subgroup augmentation of a crystal structure and pre-relaxation with an SO(3) local environment descriptor

    npj Computational Materials 2026 Journal article

    AI-Assisted Rapid Crystal Structure Generation Towards a Target Local Environment

    Osman Goni Ridwan, Sylvain Pitié, Monish Soundar Raj, Dong Dai, Gilles Frapper, Hongfei Xue, Qiang Zhu

    4 citations 2026

    Crystal structure prediction usually means expensive energy minimization over thousands of candidates, and AI generators have only handled a few tens of atoms per cell. LEGO-xtal is a symmetry-informed generator: it trains on a subgroup-augmented dataset, proposes structures, then relaxes them against local geometry descriptors instead of energy. Starting from 25 known low-energy sp2 carbon allotropes it found over 1,700, all within 0.5 eV/atom of graphite.

    Generative Models Crystal Structure Prediction Space Group Symmetry Materials Science
    Paper arXiv
  • Job runtime, arrival interval, and resource allocation distributions across the Blue Waters, Mira, Theta, Philly, and Helios clusters

    IEEE IPDPS 2024 Conference paper

    Cross-System Analysis of Job Characterization and Scheduling in Large-Scale Computing Clusters

    Di Zhang, Monish Soundar Raj, Bing Xie, Sheng Di, Dong Dai

    6 citations 2024

    HPC clusters now run deep learning jobs next to classic simulations, and the old assumptions about their workloads no longer hold. We compared job traces across five systems, Mira and Theta (classic HPC), Philly and Helios (deep learning), and Blue Waters (hybrid), looking at job geometry, arrival patterns, resource allocation, and what each means for scheduling. Joint work with Microsoft and Argonne National Laboratory.

    HPC Job Scheduling Workload Characterization Deep Learning Clusters
    Paper Demo Code
  • Cumulative distribution plots comparing real and synthetic job traces from five generative models across four clusters

    JSSPP 2024 · Springer LNCS Workshop paper, IPDPS 2024

    An Empirical Study of Machine Learning-Based Synthetic Job Trace Generation Methods

    Monish Soundar Raj, Thomas MacDougall, Di Zhang, Dong Dai

    First author 2024

    Public job traces are scarce and computing centers rarely release new ones, so synthetic traces matter for scheduling research. We compared five generative models (GAN, CTGAN, TVAE, Gaussian Copula, Copula GAN) against the manual statistical method that had been the default, judging them on distribution fit, statistical metrics, and how a scheduler behaves on the generated traces. Several of the learned generators produce usable traces with no manual tuning.

    GAN CTGAN TVAE Gaussian Copula Synthetic Data
    Paper DOI Code
  • AFDP system architecture: browser layer, pruning engine, LLM reasoning, loop detector, and adaptive feedback loop

    COMP 790 · UNC Chapel Hill Course research paper, 2026

    Loop Detection with Reflection Prompts as the Primary Driver of Web Agent Performance: An Ablation Study on Adaptive Feedback-Driven DOM Pruning

    Monish Soundar Raj, Jefrey Bergl, Rishabhdev Potti

    Led a team of three 2026

    LLM web agents choke on page observations of 10,000 to 100,000 tokens. AFDP pairs adaptive DOM pruning with a loop detector that injects a reflection prompt the moment the agent starts repeating itself. A four-condition ablation on 168 WebArena-Verified Hard tasks found that loop recovery, not pruning, drives the gains: 14.9 points more task success, while conservative pruning cut observation tokens by 52% with no change in success.

    Python Playwright Azure OpenAI Docker WebArena
    Read Paper

Side Projects

  • Swiftlet

    Swiftlet

    Swift + Metal Runtime for Streaming MoE Models on Apple Devices

    650 37

    Runs 35B and 80B Qwen mixture-of-experts models on ordinary Apple hardware, including iPhones, by keeping only the dense core resident and streaming expert weights from the SSD on demand. The 4-bit 35B runs in about 2.6 GB of RAM and the 80B in about 4.3 GB. Ships as a Swift package, a CLI, and an OpenAI-compatible local server, with every kernel validated against mlx-lm references.

    Swift Metal Mixture of Experts Apple Silicon iOS
    Code Models
  • chatcn

    chatcn

    Open-Source Chat UI Components for React

    79 7

    Beautiful, open-source chat UI components for React: message bubbles, threads, reactions, drag-and-drop file upload, voice recording, and link previews across 4 themes and 5 layouts. Installs straight into any project through the shadcn CLI.

    TypeScript React shadcn/ui Tailwind CSS
    Code Live Demo
  • ModelSweep

    ModelSweep

    Local LLM Evaluation Workbench

    20 4

    Think "Postman for local LLM evaluation": a GUI-first workbench for models running on Ollama. Build test suites, stream evaluations across models, execute generated code in isolated Docker sandboxes, score with cloud judges and round-robin peer voting, and compare Elo ratings on interactive dashboards.

    TypeScript Next.js SQLite Docker Ollama
    Code
  • Priv AI (localLLM)

    Priv AI (localLLM)

    On-Device AI · iOS App on the App Store

    14 1 1,000+ downloads

    A privacy-first iOS app that runs LLMs entirely on-device via llama.cpp. Chat offline with models from 360M to 7B parameters, get AI health coaching from HealthKit data, and track finances from PDF statements using on-device OCR. No cloud, no accounts, no tracking.

    Swift SwiftUI llama.cpp HealthKit Ollama
    Code App Store
  • GenDM

    GenDM

    Synthetic Data Generation and Management Application

    This tool is a powerful platform for generating and managing synthetic data with advanced models and metrics to ensure data quality. Its user-friendly interface simplifies dataset handling, making it ideal for research and development.

    TypeScript Python React PyTorch ExpressJS MongoDB AWS
    Code (Soon) View
  • DiffuseCampus

    DiffuseCampus

    AI Image Generation

    An innovative image generation platform utilizing fine-tuned Stable Diffusion XL architectures to create detailed, campus-specific imagery. Tailored for marketing materials, event posters, and promotional content, it delivers visuals uniquely reflective of the UNCC campus.

    Python PyTorch LoRA Diffusers Stable Diffusion HuggingFace Gradio
    Generated Img View
  • Lumos Job Trace

    Lumos Job Trace

    Data Visualization Platform

    A comprehensive web application for visualizing and analyzing job trace cluster characteristics in high-performance computing environments.

    Python Streamlit Matplotlib Pandas Numpy Seaborn
    Code View
  • Goman Project

    The Goman Project

    Digital History Platform

    A digital archive and educational platform showcasing the Haitian Revolution, featuring interactive content and historical documentation.

    JavaScript HTML5 CSS3 Omeka-S
    View

Contact

Contact Form