Rasman M Swargo

I am a Ph.D. student in Computer Science at the University of California, Irvine, advised by Dr. Sangeetha Abdu Jyothi.

I worked as a Graduate Research Assistant under the supervision of Dr. Md Arifuzzaman at Missouri University of Science and Technology. My research focused on optimizing large-scale data movement in high-speed network environments. Highlights include:

• Adaptive Concurrency for Genomic Data Downloads: Proposed FastBioDL, a parallel downloader framing genomic data retrieval as an online optimization problem. Built a dynamic concurrency controller using gradient-descent methods, achieving up to 4× speedup over the widely used NCBI SRA Toolkit and 2.1× faster transfers than static concurrency tools on 20 Gbps testbeds.
• AI-Driven Data Transfer Optimization: Architected AutoMDT, a modular data transfer architecture employing reinforcement learning for optimizing read, network, and write concurrency. Built a lightweight simulator to enable offline RL training in ∼45 minutes (vs. multi-day online training), enabling practical deployment in production testbeds. Demonstrated up to 8× faster convergence and 68% reduction in transfer completion times compared to state-of-the-art solutions.
• Foundation Model–based Network Probing: Developed npLLM, a scalable framework leveraging IBM Granite time-series foundation models for accurate throughput prediction across heterogeneous HPC networks. Fine-tuned pre-trained models with minimal training data, reducing probing error by 70% while using only one-tenth of the data required by prior solutions.

I worked as a Software Engineer - Level 2 in the Logistics and Dispatcher team at Chaldal, a startup based in Bangladesh & California, and the biggest online grocery platform in Bangladesh. My primary goal was to achieve full automation of routing operations, eliminating the need for manual interventions. The system uses a Distributed architecture with .NET (C#/F#) for backend, React for frontend and Python for algorithm and ML services, all managed on our in-house infrastructure. I was involved in designing and building backend systems, research and development of algorithms to optimize delivery processes and also some frontend development for web and mobile applications. Below I will describe some of the projects that I have worked on.

Logistics Routing Algorithm

• Goal: The goal of this project is to assign routes of multiple shipments to a delivery person so that the delivery person can deliver all the orders in the same go. In the previous system, the routes were assigned manually by the dispatcher in charge. The manual assignment is effective for low order volumes, but as the order volume increases, the manual assignment becomes inefficient.
  
• Challenge: The main challenge of this project is that the problem is an NP-Hard problem, and the number of possible routes grows exponentially with the number of shipments. The off-the-shelf algorithms like Google OR-Tools and Gurobi Optimizer were not able to handle the problem. Also, some constraints were very specific to our business and could not be modeled using the off-the-shelf algorithms.
  
• Our Work:
  • Algorithm: We developed a new routing algorithm to fully automate the route assignment process. The algorithm takes all the shipments that need to be delivered, the location of the shipments, the time window of the shipments, the weights and volumes of the shipments, the time when each of the delivery persons will be available at the hub, the vehicles that are available, and the vehicle capacity, and outputs a list of routes. A route consists of an ordered list of shipments and the delivery person who will deliver the shipments. The main objective of the algorithm is to deliver the shipments on time and to minimize the total travel time of all the delivery persons. We used Clustering, TSP, Simulated Annealing, Tabu Search, and various neighbor-generating heuristics to achieve this.
    
  • Backend: The backend system is responsible for calling the algorithm, tracking the shipment states, such as whether it's routed or not, and the estimated delivery time if it's routed, etc. It also tracks the delivery person states, like their location and the time when they will be available at the hub. The backend system keeps track of the routes assigned to the delivery persons and the shipments assigned to the routes. It updates the states of the shipments, delivery persons, and vehicles as the delivery persons deliver the shipments.
    
• My Contribution: I was responsible for designing the algorithm, implementing it, and creating a simulator for testing the performance.
• Impact: The new routing algorithm increased the on-time delivery percentage from 60% to 85%.
• Demonstration: Here is a simplified demonstration of the algorithm. All the things shown here are dummy data. The images are from the simulator that I developed to test and visualize the algorithm results, and also to simulate a total warehouse environment.

  

  

  

  

  
  In the Fig 1, there are 10 shipments shown in black and the warehouse/hub shown in blue. There are 3 vehicles available: a cycle, a bike, and a van. The corresponding routes for the vehicles are shown in the next 3 images. The colors in the shipments represent whether the shipment is on time or not. Shipments colored in yellow are early, shipments colored in green are on time, and shipments colored in red are late.

Others

• Driver Incentive System: Developed an algorithm that takes the performance of the drivers, the predicted duration of all the deliveries in a day, and the actual duration of the deliveries that the driver completed in that day, and assigns a score to each driver. The incentive system uses the score to give incentives to the drivers. 80% of the drivers reported feeling better incentivized, and the on-time delivery percentage increased by a further 5%.
• Driver Idle Time Analysis: Used Superset and Snowflake to analyze the idle time of the drivers and find the reasons for the idle time. Suggested new shift times for each driver, reducing overtime hours by 10% and minimizing drivers' idle time.
• Dish Rating and Review System: I worked for two months on the food delivery team, where I implemented a dish rating and review system. This system allows customers to rate and review the dishes they have ordered.

Published Papers

Self-Tuning Segmented Retrieval of Large-Scale Genomic Data

Authors: Rasman Mubtasim Swargo, Nieb Hasan Neom, Dr. Md Arifuzzaman (Supervisor)

Public genomic repositories such as the NCBI Sequence Read Archive and the European Nucleotide Archive are central to modern biology, but bulk data retrieval is still a bottleneck: existing tools usually fix concurrency before execution, even though throughput changes with repository load, network dynamics, and storage behavior. We present FastBioDL, an adaptive segmented downloader that uses byte-range requests, schedules segments across download workers, and uses a utility-guided gradient controller to adjust the active worker count at runtime, while segment-level progress tracking lets interrupted transfers resume without restarting completed byte ranges. Across public SRA workloads on Expanse, FABRIC, and a laboratory server, FastBioDL achieves the highest throughput in eight of nine settings, with gains up to 3.88× over the best-performing genomic-download baseline, and 1.91× the throughput of aria2c in multi-gigabit wide-area experiments.

2026 IEEE International Conference on eScience (eScience 2026), to appear. [arXiv] [Code]

Modular Architecture for High-Performance and Low Overhead Data Transfers

Authors: Rasman Mubtasim Swargo, Dr. Engin Arslan, Dr. Md Arifuzzaman (Supervisor)

High-performance scientific applications often require transferring massive datasets across wide-area networks, where static tuning of concurrency frequently leads to bottlenecks and wasted resources. To address this challenge, we developed AutoMDT, a modular data transfer architecture that leverages deep reinforcement learning to optimize read, network, and write concurrency in real time. AutoMDT employs a lightweight simulator that enables efficient offline training of Proximal Policy Optimization (PPO) agents in under an hour, eliminating the impracticality of multi-day online training on production testbeds. By decoupling I/O and network operations, AutoMDT learns fine-grained buffer dynamics and adapts to evolving conditions, achieving up to 8× faster convergence and a 68% reduction in transfer completion times compared to state-of-the-art solutions.

Published at: SC '25 Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis, pp. 939–948. [arXiv] [Code]

Deploy-Efficient and Fast Network Probing with Time-Series Foundation Models

Authors: Rasman Mubtasim Swargo, Dr. Md Arifuzzaman (Supervisor)

Monitoring network performance efficiently is crucial for high-speed networks used in modern scientific research. Traditional methods for estimating data transfer speeds are either slow or require extensive data for training, making them inefficient for dynamic networks. To address these challenges, we developed npLLM, a framework that uses time-series foundation models to predict network throughput more accurately and with minimal data, reducing probing error by 70% while using only one-tenth of the data required by prior solutions.

Published at: 2024 IEEE International Conference on Big Data (BigData), pp. 8846–8848. [Link]

Exploring Attention Mechanisms in Integration of Multi-Modal Information for Sign Language Recognition and Translation

Authors: Zaber Ibn Abdul Hakim, Rasman Mubtasim Swargo, Dr. Muhammad Abdullah Adnan (Supervisor)

Sign language recognition and translation are challenging tasks due to the complexity of body movements and expressions involved. To address these, we explored lightweight cross-attention mechanisms that enhance the integration of visual (RGB) and motion (optical flow) data. Unlike traditional methods that are computationally heavy or merge features inefficiently, our approach uses a simple plugin module that improves the understanding of movement without adding significant computational cost. It reduced the Word Error Rate (WER) by 0.9 on the recognition task and increased BLEU-4 by 0.8 on the translation task on the RWTH-PHOENIX-2014 benchmark.

Published at: 2024 IEEE International Conference on Image Processing (ICIP), pp. 2529–2535. [Link]

Manuscripts Under Review

Elastic Data Transfer Optimization with Hybrid Reinforcement Learning

Authors: Rasman Mubtasim Swargo, Dr. Md Arifuzzaman (Supervisor)

Modern scientific data acquisition generates petabytes of data that must be transferred to geographically distant computing clusters. Conventional tools either rely on preconfigured sessions that are difficult to tune without domain expertise, or adaptively optimize only concurrency while ignoring other important parameters. We present EDT, an adaptive data transfer method that jointly tunes multiple parameters: heuristic-based parallelism, infinite pipelining, and a deep reinforcement learning based concurrency optimizer. A lightweight network simulator makes agent training practical, cutting training time to under four minutes (a 2750× speedup over online training). EDT consistently outperforms existing methods across diverse datasets, achieving up to 9.5× higher throughput than state-of-the-art solutions.

Manuscript under review, 2026. [arXiv]