Kanishk Goel
ML Systems Researcher
About
I work on the systems side of large scale ML, focusing on making LLM inference faster, more efficient, and more predictable under real world workloads.
At Microsoft Research India, working with Jayashree Mohan and Ramachandran Ramjee, I study how to push the limits of serving efficiency through better request scheduling, parallelism strategies, and GPU level optimizations. A lot of my recent work explores how agent style workloads stress existing systems assumptions, and what needs to change to support them well.
More broadly, I am interested in the full stack of AI infrastructure, from kernel level optimizations to RL based post training methods that improve compute efficiency. I graduated from IIT Delhi in 2024 in Electrical Engineering with a minor in Computer Science.
Publications & Preprints
QoServe: QoS-Aware LLM Inference Scheduling
📄 ASPLOS 2026 (full paper, 14.5% acceptance rate, 152/1048) | 🪧 SOSP 2025 (poster)
A QoS-aware scheduler for LLM inference that dynamically adapts request scheduling to achieve 32% higher throughput and 10× fewer SLO violations compared to state-of-the-art baselines.
Sutradhara: System Optimizations for Agentic Inference
📄 ASPLOS 2027 (full paper) | 🪧 EuroSys 2026 (poster)
Optimizing multi-turn agentic inference by overlapping system-prompt processing with tool execution, and scheduling for maximum intra- and inter-request KV cache reuse.
MMA-Net: Multi-Modal Radar–Lidar–Camera Fusion for Autonomous Driving
📄 ICASSP 2025
A transformer-based Bird's-Eye View fusion model integrating radar, lidar, and camera modalities for robust 3D object detection in autonomous driving scenarios.
Education
B.Tech in Electrical Engineering + Minor in Computer Science
Indian Institute of Technology Delhi
2020 – 2024