CV

Research Fellow

• CoPilot Agents
  • Leading the development of LLM-powered agents aimed at eliminating technical debt in large-scale codebases by automating software development tasks.
  • Designing and implementing a dynamic AgentBuilder framework that generates agents based on customizable recipes, facilitating scalable automation of routine and complex coding operations.
• Programming with Representation (PwR)
  • This project focuses on conversational programming, enabling developers to write software using natural language instructions, decoupling the need for extensive domain and coding expertise, making software development accessible to a broader audience, including organizations and individuals without technical backgrounds.
  • Developed a system of conversational programming where natural language conversations are translated into a custom domain-specific language with built-in guardrails using LLMs, allowing users to build software easily.
  • The DSL is then translated into Python code to run on JugalBandi-Manager, enabling seamless execution of user instructions.
  • Checkout PwR-Studio and PwR-NL2DSL.
• JugalBandi - Manager
  • Leading the development and maintainence of the open source project - JB Manager, which aims to release and manage conversation bots built using PwR for non-profit usecases.
  • JB Manager is responsible for providing multi tenant support, language and interaction medium agnostic conversation flow chatbots using finite state machines.
  • Checkout JugalBandi - Manager here.
• Multi-Turn Conversational Information Retrieval
  • Developed a conversational system to streamline the form-filling process using LLMs resulting in reducing inefficiency and user experience.
  • The system dynamically guides users through follow-up questions, automatically selecting and filling the correct form from a pool of 15,000 options, significantly reducing manual effort.

Senior Data Science Engineer

• NNFetch - In House Vector Database
  • Created a dynamic recommendation platform, enhancing product discovery by blending metadata filters and partitioning with HNSW and ANNOY indexing for lightning-fast nearest product retrieval.
  • Empowered multiple data science projects with a pivotal recall set generator, adapting effortlessly to evolving metadata and delivering remarkable throughput of 1M RPM at a peak P99 latency of 20ms.
  • Delivered adaptive clustering logic, optimizing recall score and latency by extending existing indexes, enabling precise adaptation to diverse business use cases with rapidly changing metadata.

Data Science Engineer

• Personalised Search Ranking
  • Designed and developed a real-time, user-personalised search-ranking service, using long-term and short-term user features, query relevance, product attributes and dynamic product features in an XGBoost model optimised for CTR.
  • Engineered a high-performance solution for millions of users and products, leveraging Aerospike data store, SIMD-based cosine similarity processing, and ONNX Runtime for XGBoost model inferencing, delivering seamless operation at 500K RPM with a remarkable P99 latency of 30ms.
  • Proven positive impact on business metrics through AB tests, enhancing user experience and relevance at scale.
• Recommendation System
  • Engineered a versatile framework for real-time product recommendations, leveraging embedding-based features and MCDA to optimize for relevance, personalization, and real-time business metrics, and using Determinantal Point Processes (DPP) to incorporate diversity in the recommendations.
  • Scaled the system to handle throughput of 5M RPM with a P99 latency of 40ms, achieving a 19x latency reduction and an 88% annual hardware cost savings compared to the previous system.
• Machine Learning Scaling
  • Built reusable components for scaling data science models by optimizing for latency and throughput. Maintaining compatibility by python, aligning with preferences of data scientists and easy to use with minimal code changes. Proposed solutions have been adopted to improve multiple data science services at Myntra.
  • Engineered a memory-efficient solution, consolidating embedding vectors in shared memory to reduce latency by 4.5x and reduce memory footprint by 6x. Enhanced application throughput per node by enabling more worker processes per node.
  • Introduced a high-performance approach using SIMD for calculating similarity scores and re-ranking, drastically cutting computation cycles and latency. The method handles up to 15x more traffic while improving latency.
• Image Pose Detector
  • Created an scalable online service for human pose detection from an image using a ResNet-18 model with custom layers.
  • Able to handle 27x more throughput with a 6x reduction in latency from its predecessor on a single node.

Software Engineering Intern

• Developed an encoder-decoder-based generative model for preliminary image generation using Generative Adversarial Networks (GANs).
• Developed an image refinement model for processing the generated preliminary image to incorporate garment textures, ensuring accurate top-wear product replacement while retaining pose and body shape.
• The generated images were close to the Myntra image standards.

Software Developer

• Developed a music and movie recommendation system, personalised according to the preferences of multiple users in a group setting.
• Developed the backend API server responsible for serving recommendations, user and groups management and user preferance selection.

Bachelor of Technology - Computer Science and Engineering

• Relevant Courses
  • CS 13101 - Data Structures
  • CS 14101 - Analysis of Algorithms
  • CS 14103 - Computer Organization
  • CS 14104 - Automata Theory
  • CS 15101 - Microprocessor and its application
  • CS 15103 - Database Management System
  • CS 15106 - Computer Architecture
  • CS 16102 - Compiler Construction
  • CS 16105 - Computer Networks
  • CS 16103 - Data Mining

College Bot

Cloton

Distribution

Sakshatkar

Smart Ville

Sacred Bazaar

Secure Quiz

Best Long Paper Award - Convergence 23

Annual Award 2023

Quarter Award JFM 2023

Quarter Award JAS 2023

Won Multiple accolades in several technical competitions held at different techfests

• • Winner Of Softablitz (2019, 2020) - Softablitz is the software development event organized under the annual techno-management event Avishkar at MNNIT Allahabad
  • Winner of TuxWars (2020) - TuxWars is the Linux based event organized under the annual techno-management event Avishkar at MNNIT Allahabad
  • Winner of DevJam (2019) - DevJam is the software development contest organized under Weekend Of Code by Computer Club MNNIT Allahabad

Institute Merit Scholarship

• Provided by MNNIT Allahabad to top 10% of students in the batch for excellence in academics