Research

• AI infrastructure

  Multi-tenant GPU clouds on Kubernetes, scheduling and isolation, and the cost and memory trade-offs of LLM serving.

• AI/ML for cybersecurity

  Intrusion and advanced persistent threat detection, and malware detection in virtualised cloud environments.

• Explainable and neuro-symbolic models

  Interpretable deep learning for cyber-physical systems and neural-symbolic policy enforcement.

• Distributed systems and observability

  Large-scale platform development, reliability engineering and observability at platform scale.

• Nov 2025

  Explainable Multi-Modal Deep Learning Framework for Enhancing Cyber-Physical Systems Security

  Srikanta Datta Tumkur, Jay Iyer, Sanjay Babu Kurra

  2025 IEEE 12th International Conference on Cyber Security and Cloud Computing (CSCloud), pp. 48–55. DOI: 10.1109/CSCloud66326.2025.00015

• Oct 2025

  Graph Neural Networks for Advanced Persistent Threat Detection

  Srikanta Datta Tumkur, Mahima Ganapati Pushpa, Sanjay Babu Kurra, Jayesh Yadav, Svwaroopananda Kashyap

  2025 IEEE 2nd International Conference on Cryptography, Informatics, and Cybersecurity (ICoCICs), pp. 211–217. DOI: 10.1109/ICoCICs68032.2025.11384065

• May 2025

  Optimizing Malware Detection in Virtual Cloud Environments Using Hybrid Machine Learning Approach

  Srikanta Datta Tumkur, Vishwanath Eswarakrishnan, Anant Wairagade, Muhammad Adnan Aslam, Muhammad Bilal, Adnan Mustafa Cheema

  Arabian Journal for Science and Engineering (Springer), vol. 50, no. 23, pp. 19463–19494. DOI: 10.1007/s13369-025-10205-x

• May 2025

  Neuro-Symbolic Approaches for Cybersecurity Policy Enforcement

  Srikanta Datta Tumkur, Vinoth Ganesh Umamageswari Asokan

  2025 5th Intelligent Cybersecurity Conference (ICSC), pp. 327–332. DOI: 10.1109/ICSC65596.2025.11139990

• Aug 2026

  More GPUs or a Smaller Cache? Tensor Parallelism versus KV Compression for Memory-Bound LLM Serving

  Srikanta Datta Tumkur, Mehar Simhadri, Anshu Bansal, Jay Iyer, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly, Raj Dandekar

  arXiv preprint, arXiv:2608.23962

LLM serving research in preparation. Links will be added as each paper is published.

• GPU data center

  OpenDeepAPM: from request to silicon causality

  In collaboration with Anand Iyer and Mohammad Ali

  One unified trace per request: 27 spans across 7 layers.

• Benchmarking

  OpenDeepBench: from request to silicon causality and workload-grounded benchmarking of LLM serving stacks

  In collaboration with Anand Iyer and Mohammad Ali

  Chat, RAG, agent, code and document-QA workloads scored end to end.

• Routing & scheduling

  Learned routing for SLO-aware LLM serving

  In collaboration with Anand Iyer

  Prefill and decode pools, goodput, and P95 time to first token.

• Speculative decoding

  Adaptive speculation under KV-cache compression

  In collaboration with Mohammad Ali

  Speculative decoding under 4-, 3- and 2-bit KV caches.

• Memory & KV cache

  Tiered KV caching across HBM, DRAM and SSD

  In collaboration with Anand Iyer

  Session KV, predictive prefetch, and promote, demote and evict policies.

• Cost–quality frontier

  Pareto Atlas: mapping the cost–quality frontier of LLM serving

  In collaboration with Mohammad Ali

  Calibrated simulation, the dominant configuration per regime, and the tensor-parallel versus KV-compression crossover.

• Jan 2026

  Secure Multi-Tenant GPU-as-a-Service on Kubernetes: Architecture, Isolation, and Reliability at Scale

  Invited talk, 3rd World Congress on Smart Computing (WCSC 2026), Bangkok, Thailand.