CV

ML Systems Engineer · Google Cloud

Technical lead for Gemini inference optimization, Sovereign Cloud disaggregated serving, and TPU v7 (Ironwood) multi-cube scaling.

• Gemini Serving & Runtime Optimization: Driving cross-stack inference performance for Gemini models across GPU and TPU fleets, targeting XLA/runtime bottleneck elimination, KV-cache memory efficiency, and end-to-end TTFT/TPOT latency.
• Sovereign Cloud Disaggregated Serving (GDC): Spearheaded the intra-node (4:4 prefill/decode split) disaggregated serving architecture for Gemini on bare-metal nodes in Google Distributed Cloud, overcoming strict air-gapped memory isolation and control-plane constraints.
• TPU v7 (Ironwood) Architecture & Interconnect Scaling: Architected the NPI microbenchmarking suite bridging internal Borg clusters with Cloud XLA/JAX runtimes; executed multi-cube scaling analyses and collective communication (AllReduce/AllGather/AllToAll) modeling that anchored multi-billion-dollar hardware evaluations with tier-1 frontier AI customers.
• Day-1 Frontier Model Enablement (Google Cloud Tech Impact Award, 2026): Built automated continuous regression pipelines enforcing Borg-Cloud runtime parity and led XLA/kernel optimizations for open-weight LLMs (Llama 3, DeepSeek) to land Day-1 hardware Sim-Ship.

Applied Scientist · AWS, Amazon

Architected LLM-powered distributed diagnostics for hyperscale data processing and active verification for cloud network security.

• Hyperscale Spark Auto-Debugging (Amazon Q): Architected and launched a Generative AI root-cause analysis system for Apache Spark across AWS Glue, Amazon EMR, and internal compute fleets, slashing customer Mean Time to Resolution (MTTR) from days to minutes.
• Active Cloud Network Verification: Engineered a high-throughput active testing service for AWS network security, injecting synthetic live traffic to validate Network ACLs at scale and eliminating false-positive blindspots inherent to static reachability models.
• Production Control-Plane Integration: Designed and shipped real-time diagnostic and remediation pipelines directly into AWS Glue and EMR consoles, serving thousands of enterprise analytics workloads.

Research Intern (Redmond) & Research Fellow (India) · Microsoft Research

Distributed systems, edge-cloud DNN inference scheduling, and IoT analytics (Fellow 2016–2018; Ph.D. Intern 2020–2021).

• Edge-Cloud DNN Scheduling (DeepEdge & Networking Group): Pioneered distributed scheduling and low-latency offloading protocols for deep learning inference under severe compute and WAN bandwidth constraints (published at USENIX NSDI '19; granted US Patent 10,942,767).
• FarmBeats IoT Platform (Microsoft Outstanding Technical Achievement Award): Co-designed and built an end-to-end RF/sensor and cloud analytics platform for data-driven agriculture; productized into Microsoft Azure, featured on GatesNotes, and granted US Patent 11,327,476.
• Research Mentorship: Mentored 4+ research interns and junior engineers across systems and theoretical ML learnability initiatives, resulting in multiple top-tier publications.

Ph.D. in Electrical & Computer Engineering

Carnegie Mellon University

CIT Dean's Fellow

• Thesis: Practical Network Layer Machine Learning for IoT Security
• Advisors: Prof. Vyas Sekar & Prof. Anthony Rowe

B.Tech and M.S. by Research in Computer Science

International Institute of Information Technology (IIIT)

Dean's List for Academic and Research Excellence

Google Cloud Tech Impact Award

Google Cloud

Recognized for TPU v7 (Ironwood) collective scaling baselines, Borg-Cloud runtime parity, and Day-1 frontier LLM enablement (Sim-Ship).

Microsoft Outstanding Technical Achievement Award

Microsoft

Recognized for co-designing and building FarmBeats, productized into Microsoft Azure and featured by Bill Gates.

CIT Dean's Fellowship

Carnegie Mellon University

Presidential/Dean's doctoral fellowship awarded by the College of Engineering at Carnegie Mellon University.

Lumen: A Framework for Developing and Evaluating ML-Based IoT Network Anomaly Detection

ACM CoNEXT

Modular ML evaluation framework across 100+ devices, isolating and resolving temporal and topological domain-shift degradation in network traffic anomaly detection.

Lumos: Identifying and Localizing Diverse Hidden IoT Devices in an Unfamiliar Environment

USENIX Security

Mobile wireless sensing system leveraging 802.11 PHY/MAC features and AR to localize hidden IoT devices to 1.5m accuracy with 95% precision.

Contention-Aware Performance Prediction for Virtualized Network Functions

ACM SIGCOMM

Low-level subsystem contention modeling (cache, memory bus, packet I/O) to predict throughput degradation and guarantee QoS SLOs in multi-tenant cloud NFV.

All that GLITTERs: Low-Power Spoof-Resilient Light Anchors for Augmented Reality

ACM IPSN

Optical retro-reflective sensing pipeline enabling sub-centimeter AR localization at milliwatt power consumption.

DeepEdge: A Network Edge for Deep Learning Workloads

USENIX NSDI

Distributed edge-cloud scheduling framework optimizing multi-tenant DNN inference pipelines under tight compute and uplink bandwidth constraints.

Deep neural network workload scheduling

US Patent 10,942,767

Scheduling architecture for optimizing deep learning inference workloads across heterogeneous edge and cloud compute under resource constraints.

Sensor fall-curve identification

US Patent 11,327,476 (US20200150640A1)

Time-of-flight and RF signal fall-curve analytics for agricultural sensing and IoT infrastructure (Microsoft FarmBeats).

Lumos: Identifying and Localizing Diverse Hidden IoT Devices in an Unfamiliar Environment

USENIX Security • Boston, MA

Lumen: ML-Based IoT Network Anomaly Detection

CyLab & CONIX Annual Review • CMU, Pittsburgh, PA

All that GLITTERs: Light Anchors for Augmented Reality

ACM IPSN / CPS-IoT Week • Sydney, Australia

FarmBeats: AI, Edge & IoT for Agriculture

Microsoft Techfest • Redmond, WA