Arjun Srinivasan

Arjun Srinivasan

Research Engineer for Generative AI, Multimodal Learning, and Scalable ML Systems

I build research-grade ML systems that connect careful experimentation with deployable engineering. My work spans generative AI, multimodal learning, efficient training, and production inference, with experience across NSF- and DoD-funded research as well as real-world manufacturing and geospatial AI systems. I care most about reproducible experiments, strong empirical evidence, and turning good research ideas into reliable code.

Master's in Robotics Engineering University of Maryland, College Park
Bachelor's in Mechanical Engineering NIT Tiruchirappalli, India
Selected Work Research Notes Get In Touch

Selected Research & Engineering Work

OpenAI Parameter Golf: GDN Hybrid Long-Context LM

Efficient Long-Context Training
  • Replaced 7 of 9 full-attention layers with Gated DeltaNet mixers while matching the parameter budget
  • Showed a long-context crossover: hybrid loses at 8k, then wins at 16k and 32k under the same wall-clock budget
  • Improved validation bpb from 1.7059 to 1.4709 at 32k with 1.62x faster steps
  • Released configs, logs, scripts, and analysis for open-source review
Writeup PR

Qwen3-1.7B Distributed Training Throughput Optimization

Multi-GPU Systems Performance
  • Compared FSDP2, activation checkpointing, mixed precision, torch compile, and optimizer-state sharding on 2x RTX 3090s
  • Reached roughly 50% estimated MFU after tuning batch size, sequence length, checkpointing, and kernel-level efficiency
  • Profiled utilization, memory pressure, communication overhead, and step time to identify the highest-throughput setup
  • Documented a practical recipe for constrained-GPU large-model experimentation
Codebase

AI-Assisted Laser Welding System

Real-Time Multimodal Inference
  • Built a real-time inference pipeline with 3 ms latency for deployment-constrained industrial settings
  • Integrated multi-sensor fusion across camera, microphone, and 3D scanning signals
  • Reduced weld defects by 30% using ensemble deep learning
  • Deployed inference services through Triton for production use
Project Link

Trust-Aware Robot Navigation

Embodied AI & Reinforcement Learning
  • Developed socially intelligent navigation using reinforcement learning
  • Integrated natural language instructions and human intent signals into the policy setup
  • Implemented PPO in PyTorch for trust-aware policy optimization
  • Published findings at IEEE ICRA conference
Codebase

Adversarial Reinforcement Learning for Autonomous Driving

Robustness, Simulation, and Multi-Agent RL
  • Engineered a learning-based adversarial framework to generate failure scenarios for rule-based driving agents
  • Formulated adversarial rewards that reduced ego-agent cumulative reward by roughly 30%
  • Designed a defense pipeline to improve resilience under adversarial conditions
  • Extended Highway-Env for multi-agent adversarial training with kinematic observations
Codebase

AI-Driven CVD Diamond Manufacturing

NSF-Funded Applied ML Research
  • Developed hybrid CNN-LSTM models for crystal growth prediction
  • Achieved roughly 99% accuracy in forecasting growth trajectories
  • Enabled anomaly detection up to 6 hours in advance
  • Supported published work on industrial prediction and monitoring
Paper

Professional Experience

Fraunhofer USA Center Mid-Atlantic CMA
Research Data Scientist
October 2020 - Present | Riverdale, MD

Led applied ML systems work across training infrastructure, synthetic data, predictive modeling, and deployment for manufacturing and imaging systems. The work consistently emphasized reproducible experimentation, throughput-aware engineering, and production relevance.

  • Built PyTorch-based training and inference workflows with strong attention to reproducibility and operational reliability
  • Led cross-functional engineering efforts building real-time inference pipelines with 3 ms latency
  • Developed synthetic data workflows with diffusion models and GANs for manufacturing and geospatial domains
  • Published peer-reviewed work spanning robotics, industrial AI, and scientific imaging
  • Mentored junior engineers and interns in sprint planning and deployment best practices
University of Maryland, College Park
Research Assistant
June 2020 - October 2020 | College Park, MD

Developed socially intelligent navigation policy for robots using reinforcement learning, integrating natural language instructions and human intent signals for effective human-robot collaboration.

  • Designed trust-aware policy architecture using Proximal Policy Optimization (PPO)
  • Collected and curated dataset of 200+ natural language instructions
  • Simulated robot navigation in 3D environments using Unity3D and ROS
  • Published findings at IEEE ICRA conference
SimInsights
Machine Learning Intern
May 2020 - June 2020 | Irvine, CA

Built synthetic data pipeline using simulation to accelerate training and deployment of object detection models for industrial tools on edge devices.

  • Created large-scale synthetic datasets using NVIDIA Isaac Sim with domain randomization
  • Automated simulation scenes via Unity3D and Python scripting
  • Trained YOLOv3 models using NVIDIA Transfer Learning Toolkit
  • Applied TensorRT optimization for real-time inference on Jetson Nano
Fiat Chrysler Automobiles
Research and Development Engineer
July 2017 - July 2019 | Chennai, India

Designed and implemented AI-driven system to optimize in-cabin climate control in passenger vehicles using vision-based occupant detection and pose estimation.

  • Developed computer vision system using TensorFlow for passenger detection
  • Integrated real-time inference with Raspberry Pi and motor controllers
  • Collected and annotated facial orientation data across varying conditions
  • Conducted field validation testing for thermal comfort optimization

Publications

AI-Guided Feature Segmentation Techniques to Model Features from Single Crystal Diamond Growth
R. R. Mekala, E. Garratt, M. Muehle, A. Srinivasan, A. Porter, and M. Lindvall
🔗 Key Engineering Materials, 993, 67-74, 2024
AI-Guided Frame Prediction Techniques to Model Single Crystal Diamond Growth
R. R. Mekala, E. Garratt, M. Muehle, A. Srinivasan, A. Porter, and M. Lindvall
🔗 Journal of Vacuum Science & Technology A, 43(3), 032708, 2025
Can a Robot Trust You?: A DRL-Based Approach to Trust-Driven Human-Guided Navigation
V. S. Dorbala, A. Srinivasan, and A. Bera
🔗 2021 IEEE International Conference on Robotics and Automation (ICRA), 2021, pp. 3538–3545
Theoretical and Experimental Investigations on the Effect of Overlap and Offset on the Design of a Novel Quadrotor Configuration, VOOPS
G. Nandakumar, A. Srinivasan, and A. Thondiyath
🔗 Journal of Intelligent & Robotic Systems, 92, 615-628, 2018

Technical Focus

Training

PyTorch, distributed training, mixed precision, checkpointing, throughput tuning.

Evaluation

Benchmarks, ablations, reproducible experiments, and technical analysis.

Systems

Docker, Triton, ONNX, experiment orchestration, and deployment pipelines.

Tooling

Transformers, TRL, TorchTitan, vLLM, W&B, MLflow, Ray, DVC.

Writing & Research Notes

When Does It Pay Off to Swap Attention for GDN? A 600-Second Experiment
Arjun Srinivasan
📝 Long-context LLMs, efficient training, and empirical analysis
Is PCA Just SVD? A Geometric and Optimization View from ML to Deep Learning
Arjun Srinivasan
📝 Blog - Machine Learning & Linear Algebra
Demystifying PyTorch Distributed Data Parallel (DDP): An Inside Look
Arjun Srinivasan
🔗 Medium - Technical Deep Dive

Let's Connect

Let's Connect

Email

arjunsrinivasan.a@gmail.com

LinkedIn

linkedin.com/in/arjunsrinivasana/

GitHub

github.com/aarjunsrinivasan