San Francisco Bay Area
Check out: https://arbabarshad.com/ Actively seeking job opportunities - No Sponsorship Required. About me: I am a Ph.D. candidate in Computer Science at Iowa State University, working as a Research Assistant at Iowa state university. Recently I have also interned at Intel, applying Computer Vision and Large Language Models to UI automation, graphics validation, and hardware enablement. My work interests lie at the intersection of Large Language Models, Computer Vision, and MLOps.
◦ Agricultural LLM: Developed Large Language Model (LLM) for precise agricultural recommendations using expert-verified data on 90 species. Implemented Retrieval Augmented Generation (RAG) with 82% recall and 65% precision. Integrated multi-LLM support for enhanced performance. (AgLLMs.github.io) ◦ LLM Evaluation: Developed 12-task agricultural benchmark to evaluate multimodal LLMs (Claude, GPT-4, Gemini, and LLaVA). Established baseline metrics, achieving up to 73.37% F1 score with few-shot learning. (AgLLMs.github.io/AgEval) ◦ 3D Plant Modeling: Evaluated Neural Radiance Fields (NeRFs) for detailed 3D plant reconstruction in various environments. Achieved 74.6% accuracy in challenging outdoor scenarios, demonstrating NeRFs’ potential for complex modeling. Developed optimization technique reducing training time by 50% with minimal accuracy loss of 7.4%.
Applying Computer Vision and Large Language Models to UI automation, graphics validation, and hardware enablement.
Deployed auto-scaling in AWS Fargate; stress-tested API to validate container duplication and optimized resource usage. • Constructed end-to-end pipeline for routine stress tests, utilizing JMeter for scripting and Blazemeter via Taurus for cloud execution. • Customized GitLab CI/CD pipeline to execute tests seamlessly, guaranteeing no disruption to AWS resources or other development work • Received formal recognition in two sprint retrospectives for establishing the baseline for comprehensive load tests.
Contributed to the execution of 5 automated program repair tools for an empirical study on SLURM-based GPU clusters • Reduced execution time by 16x by enabling parallel execution of tools on 40 GPU clusters • Publication received a Distinguished Paper Award at the 38th IEEE/ACM International Conference on Automated Software Engineering.
• Developed 12 ML models (including Artificial Neural Network, Deep Belief Network, Random Forest, and Light Gradient Boosting) using 50 million data records of 18 features to predict monthly electricity consumption in Dubai. • Assessed model performance through 10-fold cross-validation, resulting in R2 scores ranging from 64.2% to 92.5%. • Optimized training time from 420.4 ms to 45.2 ms by using the decision tree model. • Authored and published research paper (publication) and facilitated a team of 6 data scientists in analyzing model outcomes.