New York City Metropolitan Area
I am a Quantitative Researcher in the news team at Two Sigma, working on developing trading signals and strategies using text data and LLMs. Prior to this, I was a PhD student in mathematics and machine learning, broadly working on sequential decision-making under partial observability. My research during my PhD ranged across reinforcement learning, bandits, RLHF, LLMs, LLM agents, causal inference, model predictive control and graph neural networks. I enjoy designing and implementing algorithms and methods that have strong real-world performance and are rooted in structured, generalizable ideas. Overall, my aim is to design and deploy principled methods for building AI solutions and agents that can assist with and/or improve upon human-level decision making in a safe and reliable fashion.
Working on the news team, developing trading signals and strategies using text data and LLMs.
Worked on projects in offline reinforcement learning, causal inference, double descent and denoising, using RL to improve GNN performance, latent bandits, multi-fidelity feedback, MDPs with partially observed rewards, etc Mentors and Collaborators: Ambuj Tewari (Michigan), Rishi Sonthalia (UCLA), Eli Meirom (NVIDIA), Yonathan Efroni (Meta), Aadirupa Saha (Apple), Nadav Merlis (ENSEA), Kevin Tan (Wharton), Aldo Pacchiano (Broad Institute), Kashvi Srivastava (Michigan).
Sole instructor for a section of 15-20 students in Calculus 1 or 2 (Math 115/116) every semester.
- Developed methods to use text data to predict returns, tackling overfitting and memorization. Increased proof-of-concept R^2 by ~100%. - Designed, trained and researched foundation models for order data, discovering a consistent hierarchy in performance across architectures. - Implemented both manual and automatic trading strategies in simulated markets, tackling real issues like sim-to-prod alignment, outlier events, etc.
Researched memory and retrieval mechanisms for LLMs under Adith Swaminathan and Nathan Kallus. - Designed a modular agentic memory workflow that can be plugged into any LLM agent for any task without manual-adaptation. Paper in preparation. - Demonstrated 5-10 percent gains in F-1 scores, accuracy and coverage over baselines across question-answering and code generation tasks, using only generic prompts. - Researched an optimizer workflow that can optimize parameters and prompts of a workflow based on its past runs, similar to the Trace workflow.
Working in the autobidding and ad monetization team under Ajith Moparthi. • Designed and implemented a novel algorithm for automated bidding in advertising auctions with hourly model updates, reducing latency from four days to one hour with a 44.5% error reduction. Paper in preparation. • Derived and documented a novel exact inference rule for multi‑task Gaussian processes using only dot product samples. • Wrote ∼2000 lines of code to be shipped to the team, with applications to three different team projects.