Adarsh D.

AI Researcher

Bengaluru, Karnataka, India2 yrs 3 mos experience

Most Likely To SwitchAI Enabled

Key Highlights

Optimized PyTorch for Intel's AI accelerators.
Achieved 93.8% GPU utilization in distributed training.
Designed algorithms for computational efficiency.

Stackforce AI infers this person is a Data Science and AI specialist with a focus on high-performance computing.

Contact

Skills

Core Skills

PytorchMachine LearningData Analysis

Other Skills

AcceleratorDistributed AlgorithmsDeepSpeed ZeROFSDPAlgorithm DesignLogistic RegressionSQLIT Business AnalysisPython (Programming Language)DashboardsStatistical Data AnalysisGoogle AnalyticsPredictive AnalyticsPredictive ModelingPrediction

About

As an AI Frameworks Intern at Intel, I contribute to the development and optimization of the PyTorch Framework bridge software stack for Intel’s AI accelerators and GPUs. My work focuses on building a robust, performant, and flexible software stack to support training and inference workloads, including large language models (LLMs). Collaborating with the AI Frameworks team, I leverage my expertise in machine learning and computing to enhance the performance and scalability of AI-driven solutions. This internship aligns with my passion for advancing innovative technologies in high-impact domains like AI, healthcare, and space. Pursuing an M.Tech in Machine Learning and Computing at the Indian Institute of Space Science and Technology (ISRO), Trivandrum, I bring a strong foundation in mathematics (MSc) and big data analytics (Postgraduate Diploma). My projects, including Process Mining for Bank Loans (20% faster processing), Cancer Survival Prediction, and Market Basket Analysis, reflect my commitment to solving real-world challenges through data-driven approaches.

Experience

2 yrs 3 mos

Total Experience

7 mos

Average Tenure

11 mos

Current Experience

Intel corporation

AI Frameworks Intern

Jun 2025 – Present · 11 mos · Bengaluru, Karnataka, India · On-site

Conducted extensive testing and validation of TorchFT (PyTorch Fault Tolerance) on Intel XPU architecture, ensuring robust distributed training capabilities and seamless fault recovery for production AI workloads.
Implemented and optimized distributed training strategies including DeepSpeed ZeRO and FSDP for scalable LLM training, achieving 93.8% GPU utilization across multi-GPU configurations.
Optimized compile-time performance for CUTLASS infrastructure’s Xe_gemm operations and maintained AI framework compatibility across Intel Xe series GPUs.
Presented technical and mathematical deep-dive on CUTE/CUTLASS tensor layout mapping and memory optimization strategies, covering HBM to SLM data movement patterns and physical memory management for GPU workloads to Cutlass Infra team
Designed Sliding Window Flash Attention algorithm with intelligent block classification system, categorizing attention blocks into fully-inside, intersecting, and fully-outside categories to enable selective computation and achieve 65% computational savings for narrow window configurations.
Developed proof-of-concept implementation for SWFA algorithm with per-position masking and numerical accuracy validation, delivering 50% training speedup while maintaining sub-microsecond precision across 19 test configurations.
Developing softmax division POC for Sliding Window Flash Attention based on Flash-Decoding techniques, implementing parallel softmax computation across attention blocks to reduce sequential dependencies and improve inference throughput for long-sequence processing

PytorchAcceleratorMachine Learning