Shubhankar Joshi

• Architected high-throughput data processing systems by automating Nextflow and SLURM workflows to scale raw data pipelines for millions of samples • Boosted model training efficiency by 3× on DGX-H100 GPUs through PyTorch Lightning, dynamic token batching, and optimized loss functions, enabling training on multi-billion token datasets. • Reduced GPU memory footprint by 20% while maintaining accuracy by integrating FlashAttention and RaggedMmap memory mapping, improving scalability for enterprise-scale AI workloads. • Classified 45k+ complex sequences across 160+ target categories by deploying transformer-based language models and hybrid deep learning architectures

• Achieved a Pearson correlation of 0.837 in ribosomal sequence prediction by designing and deploying a hybrid LLM-based deep learning pipeline using Transformers and CNNs in PyTorch. • Improved model performance by 10% via targeted hyperparameter optimization using Optuna and grid search, optimizing key architectural parameters across multiple training cycles. • Automated production pipelines using PyTorch Lightning and Weights & Biases, cutting manual validation time by 15% across distributed clusters • Built modular ETL and feature engineering pipelines for transcriptomic data using Python, Pandas, and Snowflake, supporting scalable processing of RNA-seq and ribo-seq datasets across multiple conditions.

• Extracted and processed 5GB of daily production data from SQL, HDFS, text files, and Excel spreadsheets. • Cleaned and preprocessed data, eliminating duplicates and blanks, using Python and SQL, resulting in a 20% improvement in data quality. • Created interactive Power Management Dashboards on Tableau, visualizing data with over 10 different KPIs, enabling data-driven business decisions. • Analyzed previous dashboards, generating actionable insights that optimized KPIs by 15%. • Collaborated in agile teams to continuously update dashboards based on user feedback, improving dashboard usability by 25%.

• Led a team of 30 members to organize online and offline music events for students, enhancing community engagement. • Managed social media campaigns on Instagram, increasing follower engagement by 30%. • Coordinated and hosted various events including open-mics, community sessions, and jamming sessions, providing a platform for over 100 students to showcase their talent.

• Reduced material tracking time 30% by integrating Python automation with factory ERP/WMS systems. • Increased OEE by 10% through real-time Power BI dashboards monitoring QC metrics for wire harness production lines. • Led 10+ shop-floor technicians to meet Piaggio delivery timelines, resolving real-time production escalations

AI 534 Machine Learning AI 539 Trustworthy ML AI 539 Optimization for AI AI 535 Deep Learning CS 578 Cyber Security AI 539 Natural Language processing CS 514 Algorithms CS 546 Networks in Computational Biology CS 558 Introduction to information Visualization