Skip to content

Sankalp Naik

M.S. Artificial Intelligence Engineering (ECE), Carnegie Mellon University

I’m a master’s student at Carnegie Mellon University and a graduate research assistant in the NEXUS Research Group, advised by Prof. Tathagata Srimani. My research spans heterogeneous memory systems, hardware–software co-design, and compiler-guided optimization for ML workloads.

Previously, I was an Associate Applications Developer at Oracle Financial Services. I received my B.Tech in Electronics & Communication Engineering from VNIT, graduating first in my class.

Passionate about intelligent systems, edge inference, and hardware-aware ML. I love turning research into production—fast, reliable, and delightful.

Pittsburgh, PA

Research

Research interests. Heterogeneous memory systems and 3D-stacked HBM; hardware–software co-design for ML workloads; accelerator design-space exploration; compiler-directed memory placement; MLIR and domain-specific languages for heterogeneous hardware; performance modeling and parallel kernel optimization.

Research vision

Modern AI systems are increasingly limited not by compute but by how well software can exploit heterogeneous memory, reconfigurable fabric, and accelerator design spaces too large to search by hand. I want to make these hardware decisions first-class, analyzable objects in the compiler: a modular, MLIR-based stack in which memory placement, accelerator configuration, and code transformation are expressed, searched, and verified within one IR, with learned models guiding the search without replacing the compiler’s reasoning. My work on compiler-visible memory (XBM), design-space search with guarantees (PROSE), and learned program representations (.ksh + TIER) builds toward this goal.

XBM: Heterogeneous 3D-stacked HBM with non-volatile base dies

First author · Manuscript in preparation · Patent pending

Designed a segmented HBM stack with FeRAM-based NVDRAM in the bottom dies and DRAM above; used the read-dominated weight and write-active KV-cache access patterns in LLM inference to motivate heterogeneous memory placement.

Research details
  • Profiled dense and mixture-of-experts LLM workloads and used PACT thermal simulation to evaluate the design. Relative to an STCO baseline, simulations showed 150% higher bandwidth on bandwidth-bound ML workloads, approximately 98% lower refresh power, and approximately 8.5 °C lower compute-die temperature.
  • Developing the architecture manuscript and an MLIR dialect for compiler-directed placement across heterogeneous memory domains; the compiler extension is ongoing.

PROSE: Region-based probabilistic design-space exploration for HLS

First author · Manuscript in preparation

Built a design-space exploration framework around the observation that Pareto-optimal HLS designs concentrate in a few axis-aligned regions; uses concentration-inequality bounds to certify regions of configurations.

Research details
  • Evaluated PolyBench kernels on three resource-constrained FPGAs against ScaleHLS-HIDA and AutoDSE; pruned approximately 90% of the configuration space while preserving near-optimal designs.
  • Extended the framework to multi-kernel workloads by composing per-kernel region sets; implemented parallel exploration and cost-model-guided autotuning to accelerate the search.

.ksh + TIER: Multi-level MLIR representations for code generation

Ongoing research

Designing Kernel Shorthand (.ksh), a compact DSL prototyped on xDSL, for LLM-generated programs that lower to optimized GPU kernels.

Research details
  • Built a C-to-MLIR analysis pipeline through Polygeist, extracting typed dataflow and control-flow graphs at multiple dialect levels; training contrastive RGCN and GraphSAGE embeddings per level with a learned merge.
  • Evaluating the embeddings as dense reinforcement-learning rewards for LLM code generation in place of binary pass/fail feedback; experiments remain ongoing.

Patent pending: “XBM: Heterogeneous 3D-stacked high-bandwidth memory with non-volatile base dies.”

Publications

Youtube Universe of Comments: A Machine Learning Approach for Systematic Classification of YouTube Comments on Custom Prepared Dataset

Sankalp Naik, A. Katre. IEEE World Conference on Communication & Computing (WCONF), 2023. Published.

Abstract

Prepared a Custom dataset of Hinglish YouTube Comments by web scraping & designing a comprehensive annotation scheme, considering various dimensions such as sentiment, topic, toxicity, and engagement. Leveraged robust ML classification models and NLP techniques to accurately capture the sentiment of Hinglish comments and identify abusive comments. Introduced the concept of Gravity of comments which will act as a reward mechanism for comments based on the computed sentiment and help YouTube reorder the display order of comments.

Machine Learning · NLP · YouTube Comments · Sentiment Analysis · Custom Dataset

[Paper]

ARMPC - ARIMA Based Model Predictive Control for Adaptive Bitrate in Streaming

Sankalp Naik, O. Khan, A. Katre, A. Keskar. International Conference on Paradigm Shifts in Communication, Embedded Systems, Machine Learning and Signal Processing (PCEMS), 2022. Published.

Best Paper Award, PCEMS 2022

Abstract

Proposed an ARIMA-based (Auto Regressive Integrated Moving Average) technique, to substitute the Harmonic Mean predictive algorithm used in the original Model Predictive Control scheme to determine the best possible bitrate for the given network conditions. Mathematically demonstrated that the proposed model provides us with improvements in the prediction of optimal bitrate for given bandwidth and buffer capacity within computational constraints of practical implementation. Analyzed and compared various methods, including deep learning techniques, to improve adaptive bitrate schemes for improved Quality of Experience (QoE) in video streaming.

Adaptive Bitrate · Video Streaming · ARIMA · Model Predictive Control · QoE

[Paper]

Experience

Graduate Research Assistant

Aug 2025 – Present

NEXUS Research Group, Carnegie Mellon University

Hardware–software co-design for processors and ML workloads, heterogeneous memory, and compiler-guided optimization. Advisor: Prof. Tathagata Srimani, Electrical & Computer Engineering.

Python, Verilog, C++, Git

Associate Applications Developer

Jul 2022 – Jun 2025

Oracle Financial Services Software (OFSS), India

Infrastructure team for a banking platform; owned performance-critical components. Optimized API latency and designed sharded databases and fault-tolerant microservices with asynchronous messaging across multi-cluster deployments.

PaceSetter Award for exceptional performance at Oracle.

Java, JAX-RS, SQL, Distributed Systems, JPA

Education

Carnegie Mellon University

2025 – Expected Dec 2026

M.S. Artificial Intelligence Engineering (ECE) · Pittsburgh, PA

GPA: 4.0/4.0. Interests: hardware-aware NAS, edge inference, systems for ML, deep learning, and robotics.

Selected coursework: How to Write Fast Code II (OpenMP/CUDA); Foundations of Computer Systems; Machine Learning; Deep Learning.

Visvesvaraya National Institute of Technology (VNIT)

2018 – May 2022

B.Tech in Electronics & Communication Engineering · Nagpur, India

GPA: 9.78/10.0. Rank 1/140. Visvesvaraya Medal of Excellence.

Selected coursework: Computer Architecture; Operating Systems; Numerical Methods & Probability Theory; Algorithms; Signals; VLSI; Machine Learning.

Teaching

Teaching Assistant, 18-645 How to Write Fast Code I

Aug 2026 – Present

Carnegie Mellon University

Guide students on code optimization, profiling, and benchmarking; diagnose performance bottlenecks, review assignments, and run labs and office hours.

Teaching Assistant, 18-738 Sports Technology

Jan 2026 – May 2026

Carnegie Mellon University

Guided AI and embedded-systems student projects.

Projects

Filter by topic
5 projects

CUDA Linear Algebra Library

GEMM, LU, TRSM, and TRSV kernels with register/shared-memory tiling and warp-level primitives; Nsight profiling to isolate memory-bound regimes.

C++ · CUDA · Linear Algebra · Nsight · Performance Optimization

Intent Classifier

Intent based smart navigation for complex financial systems.

Python · Flask · Named Entity Recognition

[Code]

Optimal Portfolio Optimization

Portfolio optimization system using quantitative finance techniques to maximize returns while minimizing risk.

Python · Optimization · Finance · Risk Management · Algorithms

[Code]

Fault-Tolerant Distributed System

A fault-tolerant distributed system implementing active and passive replication with failure detection and recovery components.

Active/passive replication with checkpointing and heartbeat-based failure detection.

Python · Distributed Systems · Fault Tolerance · Replication

[Code]

Codesign Optimization Framework

A co-design optimization framework for jointly optimizing system architecture and control parameters under performance and resource constraints.

Python · Co-Design · Optimization · Algorithms

[Code]

Technical Skills

Compilers & IR
MLIR, LLVM IR, Polygeist, dialect design, analysis passes, pass pipelines, IR lowering, static and dynamic analysis, cost modeling
Hardware & Architecture
Computer architecture, Vitis HLS, AMD Versal HBM, memory-hierarchy design, HBM and 3D DRAM, ARM64/Apple Silicon, hardware–software co-design, roofline analysis, performance modeling, thermal simulation (PACT), Verilog
Parallel Programming
CUDA, OpenMP, SIMD, tiling, warp-level primitives, Nsight
On-Device & Applied ML
PyTorch, JAX, Core ML, classic ML, LLM benchmarking, graph neural networks (RGCN, GraphSAGE), contrastive learning, RL for code generation, transformer inference
Performance-Critical Systems
C/C++ performance optimization, Java (JAX-RS), profiling and debugging, CMake, LLVM/MLIR
Infrastructure & Reliability
Benchmarking, observability, distributed systems fundamentals, fault tolerance
Languages
C++, C, CUDA C, Python, Java, SQL

Honors & Awards

  • Visvesvaraya Medal of Excellence (Rank 1/140), VNIT, 2022
  • Best Paper Award, PCEMS 2022
  • PaceSetter Award for exceptional performance at Oracle

Tech Reads

Introduction to Apache Kafka

2026/1/1 · 5 min read · Apache Kafka · Data Streaming · Distributed Systems

A comprehensive guide to understanding Apache Kafka, its architecture, and how it powers real-time data streaming applications.

[Read on Medium]

GitHub Contributions

Loading GitHub contributions…

View GitHub Profile

When I am not working on compilers

  • Karate (HDKI): the only setting where I break things on purpose and nobody files a bug report.
  • Badminton, squash, volleyball: anything with a net, where my reaction time finally beats my compile time.
  • Gym: the one search space I refuse to prune; every set gets evaluated.
  • Fishing: a recent hobby, and excellent training for waiting hours on a single HLS synthesis run.
  • Reading: the only input I do not try to compress into an embedding.

Get In Touch

Available for new opportunities. Connect on LinkedIn or send me a message.

Pittsburgh, PA

Opens your email app with your message.