CV

This is the curriculum vitae of Mohammed Musthafa Rafi.

General Information

Education

August 2023 - Present
Ph.D. Candidate in Computer Science
Iowa State University, Ames, IA
  • GPA: 3.93/4.00
  • Focus: Multi-Agent Systems, Large Language Models, High-Performance Computing, Machine Learning
  • Relevant coursework: Advanced Machine Learning, Deep Learning, Parallel Computing, Algorithm Design, Distributed Systems
August 2023 - May 2026
Master of Science in Artificial Intelligence
Iowa State University, Ames, IA
  • GPA: 3.92/4.00
  • Thesis: Benchmarking Tabular Foundation Models for Agricultural Yield Prediction
July 2018 - July 2022
Bachelor of Science in Electrical and Electronics Engineering
APJ Abdul Kalam Technological University, Kerala, INDIA
  • Dean's List

Experience

August 2023 - August 2024
Graduate Teaching Assistant
Iowa State University, Department of Computer Science
  • COM S 309 (Software Development Practices)
    Spring 2024, Fall 2024
    Mentored 8 assigned teams (40+ students) through full-stack development projects using Spring Boot, React, and MySQL; conducted code reviews and debugging sessions.
  • COM S 336 (Computer Graphics)
    Fall 2023
    Led weekly office hours for 60+ students on OpenGL rendering, shader programming, and ray tracing; maintained a 4.8/5.0 teaching rating.

Research / Internship

August 2024 - Present
Graduate Research Assistant
Iowa State University - Translational AI Center (IdeaLab), Ames, IA
  • CRAFT - Agentic Retrieval-Augmented Text2CAD System (OpenSCAD) - accepted at IEEE ICLAD 2026 [project page | paper | code]
    • Architected a fully traceable multi-agent LLM pipeline using LangGraph for stateful orchestration—conditional execution paths, inter-agent communication protocols, and hierarchical task decomposition across six specialized agents (Reasoner, Planner, Compiler, Verifier, Repair, Assembler)—with structured logging at each agent decision point, producing parametric OpenSCAD from natural language.
    • Integrated retrieval-augmented generation with GPT-5: FAISS vector search over 500+ domain artifacts (Fusion360 Gallery reconstructions and related assets) using L2-normalized embeddings, hybrid retrieval over meshes (OBJ/STEP/SMT) and synthetic CAD sequences; ~40% latency reduction via intelligent caching; built systematic evaluation pipelines for multi-agent reasoning quality. Achieved 75% semantic accuracy through sequence-based sketch–extrude–boolean construction.
    • Designed agent-level reward mechanisms through iterative self-repair validation: system-level compilation and verification signals propagate back to individual agents for targeted correction, improving task success via post-training refinement strategies—including OpenSCAD compilation success from 68% to 89%—while improving schema pass rate from 73% to 94%, maintaining 91% parametric dependency preservation and correct spatial reasoning for placement and clearances.
    • Implemented credit assignment for the multi-agent stack: fine-grained feedback from global task outcomes to individual agent actions; 40% reduction in cascading failures through localized error attribution and controlled ablation experiments.
    • Developed hierarchical assembly decomposition (Block → Parametric → Full Model). Validated on complex furniture and mechanical assemblies (tables, chairs, components) with 1000+ line JSON IR generation.
  • VISTA-CAD
    Follow-up to CRAFT on iteratively refining parametric CAD programs using visual feedback. Manuscript in preparation.
  • Distributed ML/LLM workflows on HPC
    Developed distributed training and inference workflows on university HPC (SLURM, Nova): multi-node GPU jobs, parallel data preprocessing, and reproducible experiment tracking with MLflow and Weights & Biases across 1000+ concurrent tasks.
  • TabPFN for Time-Series Forecasting on Industrial IoT
    Deployed the TabPFN foundation model on multivariate sensor streams across 1,300+ hierarchical units (68,200+ observations). Achieved state-of-the-art R² of 0.881 and RMSE of 7.1, outperforming traditional ML while reducing training time by ~100×. Demonstrated foundation model evaluation, benchmarking, and custom PyTorch training loops with gradient checkpointing; production stack included sliding-window inference, calibration, drift monitoring, and deployment-focused monitoring for robust performance.
  • NeRF pipeline (non-standard turntable capture)
    Built a Neural Radiance Field pipeline for a stationary-camera, rotating-object setup: custom camera pose estimation and view synthesis, trained and evaluated on ISU HPC with multi-GPU acceleration. Extended with NeRF-SOS for self-supervised object segmentation—achieving <2° reprojection error from 150 fixed-camera images, collaborative contrastive loss for zero-annotation 3D segmentation (15% IoU gain over supervised baselines), and SAM2-based masking (≈30% faster training), producing watertight meshes and per-object neural fields for CAD/robotics.
  • Slice 100k work
    To generate cad designs from code prompts
August 2026 - December 2026
AI Research Engineering Intern, Translational Research
Multiple Myeloma Research Foundation (MMRF) - Remote, Norwalk, CT
  • Building an agentic AI research copilot for the MMRF Virtual Lab platform (Gen3 data commons on AWS), translating natural-language research questions into cohort discovery and analysis over large-scale multi-omic and clinical myeloma datasets using LLMs, RAG, and platform APIs.

Publications

  • M. M. Rafi*, A. Jignasu, M. Saraeian, C. Hegde, A. Balu, A. Krishnamurthy: "CRAFT: Corrective and Robust Multi-Agent Framework for Text-to-Parametric CAD", IEEE ICLAD 2026. [project page] [paper] [code]
  • M. M. Rafi*, T. Ayanlade, B. Ganapathysubramanian, S. Sarkar, A. Krishnamurthy, C. Hegde, A. Balu: "Benchmarking Tabular Foundation Models for Agricultural Yield Prediction", Agri AI Workshop, AAAI 2026. [article]
  • M. M. Rafi*, A. Krishnamurthy, A. Balu: "Trustworthy LLM-Mediated Communication: Evaluating Information Fidelity in LLM as a Communicator (LAAC) Framework in Multiple Application Domains", IEEE DISTILL, 2025. [article]

Conference Experience

  • Oral Presentation - IEEE ICLAD 2026, Stanford University, Stanford, CA
    CRAFT: Corrective and Robust Multi-Agent Framework for Text-to-Parametric CAD [project page]
  • Poster Presentation - NVIDIA GTC 2026, San Jose, CA
    Natural Language to 3D Geometric CAD: A Multi-Agent LLM Framework for Design Synthesis Using Hierarchical Decomposition [poster]
  • Poster Presentation - AAAI 2026 Conference and AgriAI Workshop, Singapore
    Benchmarking Tabular Foundation Models for Agricultural Yield Prediction [article]
  • Oral Presentation - 2025 IEEE International Conference on Distributed Intelligence at the Tactical and Logical Edge (IEEE DISTILL), Pittsburgh, PA
    Trustworthy LLM-Mediated Communication (LAAC) [article]

Technical Skills

HPC & Distributed

  • SLURM, MPI, OpenMP, CUDA, Multi-GPU/Multi-Node Training, Job Scheduling, Parallel Data Processing

ML / AI Frameworks

  • PyTorch, TensorFlow, JAX, Hugging Face Transformers, LangChain, LangGraph

LLM Systems

  • Multi-Agent Orchestration, RAG Pipelines, Prompt Engineering, Foundation Model Fine-tuning, Post-Training

Edge AI & Embedded

  • NVIDIA Jetson Orin Nano, JetPack, llama.cpp, whisper.cpp, Piper TTS, On-Device Inference, Ollama

Languages

  • Python, C++, CUDA C/C++, Bash, SQL, Java, JavaScript, Swift

Infrastructure & Tools

  • Docker, Kubernetes, AWS, GCP, MLflow, Weights & Biases, Git, Linux, CI/CD, Unity

CAD & Engineering Software

  • OpenSCAD, SolidWorks, ANSYS, FEA, CFD

Data Analysis & Visualization

  • Pandas, NumPy, Matplotlib, Plotly, Paraview, VTK

Projects

Jarvis-home - Local AI Voice Assistant on Jetson Orin Nano

  • Built a fully on-device AI assistant with zero cloud dependency, combining whisper.cpp for speech-to-text, Gemma 4 E2B via llama.cpp for reasoning, and Piper TTS for synthesis. Designed a local intent router for lights, weather, and vision queries; actively iterating on conversation memory, on-device wake word ("Hey Jarvis"), and a multimodal vision path.
  • Jetson Orin Nano, JetPack, llama.cpp, whisper.cpp, Piper, Edge AI
  • Code | Main GIF | Highlight Image

narrAIt - Agentic Accessibility Companion for macOS

  • Co-developed an AI-powered accessibility assistant that explains on-screen elements in real time and answers voice questions in context, pointing users to exact interaction targets without taking over control.
  • Won 3rd Place at Swan Hacks Spring 2026, with a $1,200 prize and NVIDIA Jetson Orin Nano (per team member).
  • Swift, macOS, Agentic AI, Computer Use, Voice Interface
  • Code | Devpost | Demo Video | Document (PPT) | Highlight Image

VoiceForge VR - Voice-Driven Virtual Reality Interaction

  • Built an interactive Unity VR experience targeting Google Cardboard on iPhone, with FPS locomotion, gaze-based UI for stereo mode, and voice-prompt-to-STL generation in-world. Focused on intuitive in-world controls and natural interaction flow for low-cost mobile VR.
  • Unity, C#, VR, HCI, Voice Interface
  • Demo Video | Document (Final Report) | Highlight Image

Distributed 3D Gaussian Splatting & Neural Rendering on HPC

  • Investigated multi-GPU training performance for neural radiance fields and 3D Gaussian Splatting on Iowa State's Nova cluster using SLURM; benchmarked scaling efficiency across node configurations with MPI-based communication and conducted controlled experiments on rendering quality vs. compute tradeoffs.
  • SLURM, MPI, Multi-GPU, 3DGS, NeRF

LLM-Mediated Communication Framework (LAAC)

  • Developed a trustworthy communication framework addressing "communication theater" problems in multi-agent LLM workflows; designed evaluation pipelines measuring information fidelity across multiple application domains with systematic benchmarking on reasoning quality.
  • arXiv

TraitViz - RGBD Depth Estimation & 3D Reconstruction

  • Implemented state-of-the-art depth estimation and 3D reconstruction algorithms on the NYUD dataset using RGBD data from pinhole cameras; built a full pipeline from intrinsic-aware back-projection to point-cloud reconstruction and qualitative evaluation.
  • Python, OpenCV, NYUD, Point Clouds, Depth
  • Code | Video | Document

CNN Visualization using CUDA and OpenGL

  • Interactive system inspecting CNN intermediate layers with per-layer activation maps and filter weights. Implemented forward hooks and epoch snapshots for real-time training introspection.
  • PyTorch, CUDA, OpenGL, CuPy/PyCUDA, GLFW
  • Code | Video | Document

Prompt-to-Perception Pipeline

  • Built an end-to-end system with prompt refinement (T5-Small fine-tuning), Stable Diffusion 2.1 synthesis, and SAM2 segmentation. Achieved 3× ROUGE score improvement and 0.72 prompt-image alignment.
  • Code | Demo Video | Document

Diseased Plant Detection: Comparative Deep Learning Study

  • Trained and evaluated ResNet50, VGG16, and a custom CNN for plant disease classification across 38 classes on a large-scale plant disease dataset; benchmarked architectures on accuracy, training stability, and inference latency.
  • PyTorch, ResNet50, VGG16, Transfer Learning
  • Document

Certifications & Awards

Certifications

  • NVIDIA DLI: CUDA C/C++ Fundamentals for Accelerated Computing - Score: 98%
  • MLOps: Machine Learning in Practice (Digital Badge, Nov 2024, Translational AI Center, ISU) - production ML pipeline with 99.9% uptime
  • Neural Radiance Fields & Implicit Neural Representations (Digital Badge, May 2025, TrAC, ISU)
  • Complete Python With DSA Bootcamp by Krish Naik, Udemy (In progress)
  • Complete Machine Learning, NLP Bootcamp with MLOps and Deployment by Krish Naik, Udemy (In progress)

Awards

  • 3rd Place, Swan Hacks Spring 2026: narrAIt (agentic accessibility companion for macOS); received $1,200 and an NVIDIA Jetson Orin Nano (per team member)
    Hackathon | Project | Demo
  • Winner, Applied AI Challenge 2026: AI-Assisted Making Award ($1,000), for StudyBuddy - an AI-driven 3D-printable desk organizer system
    Certificate
  • Winner, Applied AI Challenge 2025: Social Impact Award ($1,500), for Karma AI - an assistive multimodal system serving 500+ users with disabilities
    Code | Demo
  • Constellation Prize (Top 4/30+ teams), Ivy Data Visualization & Storytelling Case Competition, Ivy College of Business
  • 2nd Place, Fall 2023 Coding Contest, CSE Programming Club

Contact