Ruchir Garg

Machine Learning Research Intern

• Project: Recognition of T Cells Causing Asthma Using Machine Learning Techniques.
  • Advisor: Dr. Vijayanandan Pandurangan.
  • Objective: To identify different T cells involved in Asthma through advanced ML techniques.
  • Methodology: Employed encoding techniques for the structure of T-cell and antigen and utilized an attention mechanism to analyze the interactions between T-cell and antigen.
  • GitHub Repository: ATTENTION_TCR
  • Novelty and Impact: This approach significantly enhanced the identification process of potential T cells compared to existing methods. Using this novel methodology, the number of T cells that can be tested for the VDJb database was substantially increased, demonstrating a considerable advancement over traditional techniques. The findings are currently undergoing wet lab validation to verify the new TCR-antigen binding groups.

Machine Learning Research Intern

• Primary Role: Design and build a Federated Learning infrastructure to enhance data access and cataloging.
• Key Contribution: Developed a secure framework for asynchronous model distribution to multiple clients, focusing on privacy and data security.
• Algorithm Implementation: Implemented various robust aggregation algorithms within the framework, including FedAvg, Scaffolding, and Secure Aggregation, to ensure effective and secure model training.
• Client Selection Strategy: Devised a methodology for selecting clients during model aggregation based on the number of learning cycles per client, to maintain consistency with the global minima.
• Integration with AWS SageMaker: Streamlined the entire federated learning process by integrating it with AWS SageMaker, enhancing operational efficiency.
• Research Foundation: My work aligns with the pioneering concepts in Federated Learning, as outlined in Google's research blog. For more details, see Google Research Blog on Federated Learning.

Machine Learning Research Intern

Desingn and Build Federated Learning Infra

• Project Focus: Efficient Identification of Sequencing Errors to Improve Haplotype Genome Assembly.
• Advisor: Dr. Vikas Bansal.
• Research Approach: Developed probability models leveraging neighboring reads to accurately process false variant calls.
• Methods:
  • Utilized Convolutional Neural Networks (CNN) with layers representing different genomic features to enhance accuracy in identifying in-dels in genome assembly.
  • Applied Maximum Likelihood Estimation (MLE) and Expectation-Maximization (EM) algorithms, combined with beam search, to pinpoint false variant positions and reconstruct the original sequence from noisy aligned fragments.
• Novelty and Impact: This innovative approach challenges conventional methods by exploiting the inherent similarities in human genome structures.
• Outcomes: Successfully surpassed the State of the Art (SOTA) with an accuracy of 99.7%, significantly advancing the field of genomic assembly.
• GitHub Repository: For detailed insights, visit HAPCUT_preprocessing.

Research Intern

• Project Focus: Implementation and Testing of Model Order Reduction Techniques on GPU.
• Advisors: Dr. Jens Saak (MPI), Dr. Kapil Ahuja (IIT Indore).
• Objective: To implement and validate the viability of various Model Order Reduction (IRKA) techniques for dimensionality reduction, specifically on GPU platforms.
• Methodology: Implemented algorithms in CUDA, focusing on Block QR decomposition, to optimize and test on several Linear Dynamical Systems such as HEAT BEAM models for assessing correctness and stability.
• Performance: Achieved approximately 15x amortized efficiency improvement compared to CPU-based implementations.
• IRKA Reference: For more information on IRKA, visit Wikipedia.
• Algorithm reference: Algorithms implemented can be found on the MPI-Magdeburg MORLab Project.
• Research Report: Detailed insights are available in the research report, accessible here.
• Conference Poster: The conference poster showcasing this research can be viewed here.

Engineering Intern

• Key Project: Optimization and Enhancement of the Internal Search Engine.
• Technologies Used: Elasticsearch.
• Improvements Implemented:
  • Engineered solutions for handling queries with typos and incomplete words, making the search engine more robust and user-friendly.
  • Developed case insensitivity in search queries, enhancing the search functionality for a wider range of user inputs.
  • Optimized search algorithms to return relevant results based on user ratings, page views, and user feedback, significantly improving the accuracy and relevance of search results.
• Impact on Customer Reach: These enhancements attributed to a substantial increase in customer reach and engagement, as reflected in improved user satisfaction and higher interaction rates with search results.
• Contribution to Topic Modelling: Implemented topic modelling to group related articles, which streamlined the content discovery process for users and aided in efficient information navigation.
• Company Growth: Interned during the nascent stages of GeeksforGeeks, which is now recognized as one of the largest repositories for technical blogs globally.

Quantitative Analyst

• Intraday Long-Short Equity Strategy:

  • Objective: Developed an equity-based intraday long-short strategy using institutional flow data.
  • Performance: Achieved a live Sharpe ratio of 3.7.
  • Key Technique: Leveraged order book data for strategic exits.

• Beta Neutral Sentiment Strategy:

  • Data Source: Utilized DowJonesNewsWire for sentiment analysis.
  • Approach: Combined sentiment signals from an instruct LLM model with technical signals using convex optimizaiton, and used Markowitz optimization for allocations.
  • Hedging: Employed sector ETF shorts for hedging long positions.
  • Performance: Achieved a live Sharpe ratio of 1.93.

• Event-Based Sentiment Strategy:

  • Feature: Developed sentiment features based on text and topic analysis.
  • Hedging: Strategy hedged with sector ETFs.
  • Control: Adjusted signal combination coefficients based on market regime and fundamental features.
  • Performance: Reached a Sharpe ratio of 2.33 in testing phase.
• Current Project: Working on a reinforcement learning-based allocation algorithm for the sentiment strategy.

Sr. Software Engineer (Python/ C++ Infrastructure)

• Developed advanced sorting algorithms for multi-column data in Cython C++, accommodating custom objects like NAT (Not a Date), achieving a 5x amortized efficiency gain.
• Redesigned the strategy testing overlay system from file-based to key-based, leading to a 10x improvement in data freshness and robustness:
  • Replaced initial file replication system to reduce stale data issues.
  • Implemented a lazy update mechanism for cached data in strategy classes, keyed to specific requirements.
  • Introduced diverse file-based systems (H5, Pkl, Redis backends) within the overlay for optimized storage and readability.
• Designed and maintained the Portfolio Optimization Pipeline Implemented Proximal Policy Optimization based Reinforcement Learning strategy.
• Executed optimizations for the storage and querying of extensive historical data (up to ten years), resulting in a 6-7x improvement in data retrieval times:
  • Applied sharding on custom key axes and parallel chunked updates.
  • Research and deploy repartitioning schedule of h5 for non-contiguous data, realizing (1.5x amortized efficiency)
  • Refined data alignment with non-prefix keys using almost-dense vs. sparse optimizations during cache retrieval.

Software Engineer (Python/ C++ Infrastructure)

• Engineered a library for the serialization of 'exotic' Python objects for distributed computing, achieving a 100x performance advantage over solutions like dill and cloudpickle.
• Automated the sharing of private data linked to strategy classes, enhancing team collaboration and efficiency while maintaining code confidentiality.
• Made significant modifications to OpenMP, MPI, MKL, HSL, and IPOPT libraries for uint8, uint16 object support, resulting in a 4x improvement in benchmark efficiency.