Atharva Deshmukh

Bengaluru, India

I’m a Research Fellow at Microsoft Research India, where I focus on LLM inference optimization, congestion control, and sustainable AI inference. I’m a core contributor to AI-Greenferencing — Microsoft’s effort on power-, hardware-, and workload-aware routing for renewable-energy-powered AI inference clusters.

My research focuses on LLM systems and applied ML. Recent work includes the XWind cross-site router for LLM inference at renewable-energy farms (accepted at SoCC 2026) and BeLLMan on LLM congestion control (FAISys 2025); I’ve also published at ACL 2023, with additional work under review at Neurocomputing and at Computer Speech & Language. Previously, I interned at IBM Research (LLM training efficiency) and CNeRG Lab, IIT Kharagpur (efficient fine-tuning). I graduated from IIT Patna in 2025.

Outside research, I’ve led teams to top finishes at the Inter-IIT Tech Meet and Amazon ML Challenge 2024 (3rd), founded the Research Club at IIT Patna, and headed the Machine Learning Club at NJack. I also play chess and do theatre. I love to trek (especially the Himalayas!).

Feel free to drop me a line — I’m always happy to chat about LLM systems, sustainability in AI, or research collaborations.

news

Oct 17, 2026	XWind: A Cross-site Router for LLM Inference Serving at Renewable Energy Farms has been accepted at SoCC 2026.
Apr 15, 2026	New preprint! XWind: A Cross-site Router for LLM Inference Serving at Renewable Energy Farms.
Apr 10, 2025	Our paper BeLLMan: Controlling LLM Congestion has been accepted at FAISys 2025 🎉
Jan 15, 2025	Joined Microsoft Research India as a Research Fellow — working on AI-Greenferencing (renewable-energy-aware AI inference) and LLM inference optimization 🚀
Sep 15, 2024	Secured 3rd Rank in the Amazon ML Challenge 2024 out of 1000+ teams 🥉

selected publications

SoCC

XWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms

Tella Rajashekhar Reddy^*, Atharva Deshmukh^*, Liangcheng Yu, and 7 more authors

In ACM Symposium on Cloud Computing (SoCC), 2026

Accepted

Abs Bib

@inproceedings{reddy2026xwind,
  title = {XWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms},
  author = {Reddy, Tella Rajashekhar and Deshmukh, Atharva and Yu, Liangcheng and Zhang, Chaojie and Shepperd, Mike and Gandhi, Rohan and Parayil, Anjaly and Iyengar, Srinivasan and Manchepalli, Ajay and Bhattacherjee, Debopam},
  booktitle = {ACM Symposium on Cloud Computing (SoCC)},
  year = {2026},
  note = {Accepted},
}

FAISys
BeLLMan: Controlling LLM Congestion

Tella Rajashekhar* Reddy, Atharva* Deshmukh, Karan Tandon, and 3 more authors

In Proceedings of the 1st Workshop on Foundations of AI Systems (FAISys), 2025

Abs Bib

We study congestion in LLM inference and propose token-level output reduction together with predictive scheduling to improve throughput, fairness and energy efficiency in vLLM-style serving systems.
@inproceedings{reddy2025bellman, title = {BeLLMan: Controlling LLM Congestion}, author = {Reddy, Tella Rajashekhar and Deshmukh, Atharva and Tandon, Karan and Gandhi, Rohan and Parayil, Anjaly and Bhattacherjee, Debopam}, booktitle = {Proceedings of the 1st Workshop on Foundations of AI Systems (FAISys)}, year = {2025}, }

Preprint

Dynamic Negotiation Landscapes: MBPS and the Interplay of Buyer Personalities

Subrata* Das, Atharva* Deshmukh, Sriparna Saha, and 3 more authors

2025

Under review at Neurocomputing

Abs Bib HTML

@unpublished{das2025mbps,
  title = {Dynamic Negotiation Landscapes: MBPS and the Interplay of Buyer Personalities},
  author = {Das, Subrata and Deshmukh, Atharva and Saha, Sriparna and Ramnani, Roshni and Maitra, Anutosh and Sengupta, Shubhashis},
  year = {2025},
  note = {Under review at Neurocomputing},
  url = {https://arxiv.org/pdf/2510.15330},
}

ACL
APTSumm at BioLaySumm Task 1: Biomedical Breakdown, Improving Readability by Relevancy-Based Selection

A.S.* Poornash, Atharva* Deshmukh, Archit* Sharma, and 1 more author

In Proceedings of the 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks, 2023

Abs Bib HTML

A three-step abstractive summarization pipeline that breaks biomedical articles into sections, generates candidate summaries with SIMCLS, and selects the top paragraph per section. Achieved 2nd place on readability metrics with high ROUGE.
@inproceedings{poornash2023aptsumm, title = {APTSumm at BioLaySumm Task 1: Biomedical Breakdown, Improving Readability by Relevancy-Based Selection}, author = {Poornash, A.S. and Deshmukh, Atharva and Sharma, Archit and Saha, Sriparna}, booktitle = {Proceedings of the 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks}, year = {2023}, publisher = {Association for Computational Linguistics}, url = {https://aclanthology.org/2023.bionlp-1.61.pdf}, }