Experience

Internships

• Citadel - Software Engineer Internship

  Summer 2025, Commodities, London

  • Built a robust listener and parser for trade messages, boosting accuracy via async LLM integration
  • Developed a real-time health monitor to validate curve markers, track price source heartbeats and uptime, ensure snapshots, and detect cushion/crossed states with Slack alerts across all markets
  • Conducted regression on prelim vs close prices to detect curve/source discrepancies and improve reliability
  • Extended trade upload tool with trade and risk date support for accurate tracking and downstream use

• Tower Research Capital - Quantitative Developer Internship

  Winter 2025, North Moore Trading Team, Gurugram

  • Analyzed trade data feeds received from a crypto exchange across multiple IP addresses, identifying and prioritizing better-performing connections to the firm's quote server
  • Improved connection reliability by evaluating and comparing latency and stability across exchange IPs, routing traffic through the most consistent endpoints for real-time market data
  • Integrated hardware timestamping into OpenSSL connections to achieve higher timing precision for incoming market data packets

• BharatGen - Applied Scientist Internship

  Summer 2026, Mumbai

• Amazon - Applied Scientist Internship

  Summer 2024, Bangalore

  • Worked on Amazon's Large Language Model Olympus and improved its instruction following ability
  • Implemented Classifier-free Guidance method to enhance focus on key parts of user queries and system prompts, optimizing the balance between conditional and unconditional probabilities using a hyper-parameter
  • Evaluated the performance of Olympus and some open source models on various single and multi-turn datasets

Research Experience

• Bachelor's Thesis Project - Requests of a Feather Must Flock Together: Batch Size vs. Prefix Homogeneity in LLM Inference

  2025-26, Prof. Mythili Vutukuru

  [arXiv] | [Report] | [Presentation]

  • Designed Feather, a prefix-aware LLM inference scheduler using reinforcement learning to balance batch size and prefix homogeneity, achieving 2-10x higher throughput over state-of-the-art baselines (vLLM, SGLang) integrated into both frameworks
  • Built the Chunked Hash Tree (CHT), a lightweight data structure replacing expensive radix-tree traversals for prefix detection, reducing CPU scheduling overhead by up to 1000x while keeping it under 1% of GPU execution time
  • Showed through experiments that smaller, prefix-homogeneous batches outperform larger heterogeneous ones, and that even a single divergent request can cut decode throughput by ~2×, motivating the core design of Feather

• Research and Development Project - Compressive Lognormal Regression

  2024-26, Prof. Ajit Rajwade

  • Enhanced viral load estimation in pooled RT-PCR via Bayesian inference and compressed sensing
  • Developed a superior Orthogonal Matching Pursuit (OMP) variant using combinatorial group testing
  • Optimized sparse recovery using block gradient descent, LASSO, NNLS and proximal algorithms