Hi, I’m Abhijit!
I’m an M.S. Computer Science student at UMass Amherst and a Graduate Research Assistant at Argonne National Laboratory. I have 2+ years of research and engineering experience across ML systems, distributed infrastructure, federated learning, and LLM post-training. I’m looking for full-time opportunities starting after graduation.
At Argonne, I work on large-scale LLM post-training and federated learning, building distributed training systems and tooling that support ML research across scientific domains.
I built and open-sourced HiveWatch, a distributed-training observability toolkit adopted within APPFL, and built Kubernetes-based researcher environments on NERSC Spin. Previously, I helped scale Flotilla at IISc and co-developed FedProj, published in TMLR. My work also spans confidential computing and secure federated learning.
I’m particularly interested in GPU programming, LLM serving, and scientific ML. Our work on Towards Global Federated Genome-Wide Association Meta-Analysis Using GA4GH TES received a Best Poster Award at IEEE eScience.
Experience
- Graduate Research Assistant
Argonne National Laboratory (2026-Present) - Associate Software Engineer
IUDX-IISc (2024-2025) - FL Research Intern
DREAM:LAB, IISc (2023-2024) - NLP Research Intern
Carelon (2023)
Education
- MS CS, UMass Amherst
2025-May 2027 (expected) - BTech CS, VIT
2020-2024
Interests
- GPU Programming and LLM Serving
- Distributed Training and LLM Post-Training
- Scientific ML and Federated Learning
Skills
Primary language across research projects, federated systems, and ML experiments
Built CNNs, Transformers, diffusion pipelines, flow-matching world models, and custom dataloaders for FL frameworks
PyTorch FSDP on Aurora, multi-node training on Polaris, and custom CUDA kernel development
Deployed Confidential Containers and Kubernetes-native provisioning systems (APPFL on NERSC Spin) with runtime attestation
Used heavily across research deployments and personal projects
Architected TEE-based federated learning deployments with cryptographic remote attestation
Scaled Flotilla to 1,024 clients and orchestrated large-scale FL experiments
Experiment tracking, observability (HiveWatch), and distributed training tooling across APPFL
Fine-tuned BERT, RoBERTa, and clinical language models with LoRA and prompt-based methods
Systems scripting and automation for heterogeneous cluster deployments
Deployed a full ROS autonomy stack (SLAM, AMCL, move_base) on real hardware for an autonomous delivery robot
Used for data management and experiment logging across research projects
Data preprocessing, baselines, and statistical analysis of federated learning results
Applied in LLM reasoning and agentic framework experiments
Education
Expected graduation: May 2027. Focused on ML systems, distributed training, and scientific AI.
Relevant Courses:
- Robotics
- Advanced Generative AI
- AI Alignment
- Advanced NLP
- Advanced ML
- Distributed Systems
Undergraduate degree in Computer Science with focus on machine learning, distributed systems, and software engineering.
Experience
- Develop Llama-3.1-70B and GPT-OSS-20B post-training workflows (SFT/DPO/KTO) using PyTorch FSDP, with scaling experiments up to 256 Aurora nodes (1,536 Intel GPUs); profile bottlenecks and debug out-of-memory and checkpointing failures.
- Engineered a Kubernetes-native provisioning system on NERSC Spin for APPFL, a DOE federated learning platform, dynamically spinning up org-scoped, oauth2-proxy-authenticated Jupyter environments with Keycloak SSO and an invite-based FastAPI/PostgreSQL onboarding pipeline.
- Built and open-sourced HiveWatch, a framework-agnostic observability toolkit for distributed ML training with real-time metric dashboards, geographic client visualization, and experiment tracking via WandB and MLflow, adopted across APPFL.
- Engineered federated LoRA fine-tuning on the Evo-2 model pipeline for genomic AI training across institutions, reducing cross-site data transfer to 20 MB (down 74%) on Polaris/NERSC HPC.
- Designed and implemented FedProj, a gradient-projection algorithm for federated learning that prevents catastrophic forgetting on non-IID data, achieving 2–9% accuracy gains over state-of-the-art FL baselines across 100+ clients.
- Orchestrated 70+ large-scale training runs across CV (CIFAR-10/100, CINIC-10) and NLP (MNLI, SST-2, MARC) benchmarks; work published in TMLR.
- Architected confidential computing solutions using Multi-Party Computation, Differential Privacy, and AMD SEV-SNP TEEs with cryptographic remote attestation for healthcare federated learning deployments.
- Deployed Confidential Containers (Peer Pods) on Kubernetes with runtime attestation, enabling privacy-preserving ML across multi-cloud infrastructure while reducing setup time by 60%.
- Co-developed Flotilla, a modular federated learning framework for heterogeneous resources.
- Executed 90+ experiments on heterogeneous clusters (46 Raspberry Pis, 12 Nvidia Jetsons) benchmarking IID and non-IID performance.
- Scaled Flotilla to 1,024 AWS clients, achieving 55% lower training-round latency than Flower in benchmarks by removing serialization in the communication batching path.
- Designed CPU and memory visualizations contributing to a publication in Elsevier JPDC.
- Engineered a Memory Module into ClinicalBERT, DistilBERT, RoBERTa, and Clinical-Longformer for sentiment analysis, achieving 90% accuracy.
- Transitioned centralized models to federated learning, improving accuracy by 2% and reducing training time by 30% with LoRA, P-Tuning, and Prompt Tuning.
Awards
Best Poster Award · IEEE eScience 2026
Towards Global Federated Genome-Wide Association Meta-Analysis Using GA4GH TES
Abhijit Chunduru, Matthew Joel, Zilinghan Li, and Ravi Madduri.
Selected Projects
Publications
A flexible and lightweight federated learning platform for edge environments with modular strategy support, asynchronous updates, and strong fault tolerance.