Portrait of Ansh Nagwekar

Ansh Nagwekar

LLM Research @ Susquehanna

About

Call me Ansh Nagwekar (ahnsh nah-GWAY-kar). You've just boarded my life's figurative Pequod. No white whale in sight, but I'm still looking…

Let's rewind: I grew up in sunny San Jose, CA and spent high school at Bellarmine (BCP) tinkering with robots, playing ice hockey, hiking with my Boy Scout troop, and eating an unholy number of curly fries in the cafeteria. Somewhere in there, especially during the long, remote, book-heavy stretch of COVID, I fell into the habit of seeing building as both an art and a science — a dualist little epiphany that's stuck with me ever since. That's what influenced me to major in an interdisciplinary program at Penn called NETS that fused computer science, mathematics, and economics with applications to networked systems (such as the Internet, capital markets, search engines, etc.). It eventually got succeeded by the Artificial Intelligence degree (I guess it was the first of many things to get replaced by AI).

Right now, I'm a machine learning researcher at Susquehanna, applying frontier developments in AI to our news-driven trading strategies. I joined as an intern in 2024 and went full-time in 2025 on our flagship options market-making desk. My early work there was building the systematic trading platform behind long horizon volatility-based signals: market-making strategy feeds, trading parameter harness that enable strategies to adapt intraday, and the real-time monitoring around them. Working that close to the strategy critical path changed how I think about pursuing applied research — an idea is only interesting to me if it surfaces itself in the problems we face in real time. It also reinforced the dualistic view of my work: the craft of a system and the truth of its numbers have to hold up at the same time or neither one matters.

Before that I moved through some fairly different worlds. At Ion Protocol I was a blockchain research engineer writing Go, Rust, and Solidity for DeFi lending markets and yield oracles. At Tesla, I worked on pricing and configuration software across the Tesla product lineup and supercharger network, where I used enough applications of graph traversals to perhaps compensate my algorithms professor for the hours he spent watching me fall asleep or stare blankly at the chalkboard. Before either of those, I was a founding engineer at VO2 Fans, a platform my friends and I came up with that would “tokenize” the sports fan experience. It was a pretty wild time: picture a bunch of college kids trying to create a secure Web3 platform that would transact earnings for the most engaged fans in crypto, while also trying to partner with athletes (imagine explaining NFTs to MMA stars like Tony Ferguson). We raised a $350k seed, went through Techstars' 2022 sports accelerator in Indianapolis, and eventually sold the platform away before we came back to school.

After trying out startups and blockchain, I decided to dedicate the rest of my studies to understanding and advancing the frontier of deep learning. So I enrolled in the Accelerated Master's in Computer Science and spent the extra credits researching neural network optimization with Nikolai Matni's Unstable Zeros group at Penn's ASSET Center. Some ideas that we deeply explored, such as layer-wise preconditioning, higher-order optimization algorithms, and connections to statistical learning theory, surface in our ICML 2025 paper and my graduate thesis, which was primarily written as a “cookbook” for the theory of training large neural networks at scale. Alongside that, I spent three years as an investment researcher at Open Core Ventures, where I provided technical due diligence on 250+ open-source projects, which is where most of my instincts about what good software looks like actually come from.

What ties all of this together is how quickly I pick up new technologies and product spaces — and how completely I fall into them once I do. News trading, options market-making, DeFi protocol design, EV supercharger pricing systems, optimization theory: different worlds, but I've never had trouble getting obsessive about the internals of whatever I'm in. These days that curiosity is pointed at making large models train and serve efficiently, and understanding why the optimizers we already use work as well as they do: a question that's equal parts art and science.

Back

Research

  1. ASSET Center @ Penn Engineering

    2024 – 2025

    Deep Learning Research

    Studied optimization theory and its applications to deep learning in Nikolai Matni's Unstable Zeros group. Published work relating theoretical results from statistical learning and optimization algorithm design to efficient representation learning in deep networks via layer-wise preconditioning.

  2. Graduate Coursework @ Penn Engineering

    2023 – 2025

    Academic Research Projects

    Master's-level courses where the deliverable was an original research project rather than a final exam.

    • Welfare and Click Prediction for Autobidders and Online Auctions

      Auto-bidding theory generally assumes click-through rates are known in advance; this treats CTR estimation as part of the problem and asks what that assumption costs in liquid welfare and individual regret. We developed incentive-compatible allocation algorithms for repeated single-slot ad auctions among budget-constrained bidders, covering both the myopic case, where a bidder optimises only the current round, and the non-myopic case, where it plays across all of them. The myopic algorithm reaches optimal liquid welfare over a long enough horizon, and simulations sweep a hybrid objective λ interpolating between pure value- and utility-maximisation.

      Ansh Nagwekar, Arnab Sircar, Anirudh Bharadwaj, Terrance Ji CIS 6200: Machine Learning in Game Theory

    • Spectral Dynamics of Variational Autoencoder Latent Spaces

      An empirical study on CIFAR-10 of how architectural choices shape the spectral and information-theoretic structure of a VAE's compressed representations. Standard architectures turn out to produce isotropic latent spaces, with variance spread uniformly across every dimension, while task-specific supervision measurably sharpens that structure and end-to-end classifiers learn more informative representations still. Mutual information analysis suggests why: purely reconstruction-driven training can diffuse the features that actually matter, where a task-oriented objective tends to preserve them.

      Ansh Nagwekar, Mohul Aggarwal, Dominic Olaguera, Sahishnu Hanumolu ESE 5460: Principles of Deep Learning

    • Uncertainty Estimation Techniques and Applications for Medical Insurance Cost Forecasting

      Medical cost forecasting needs calibrated uncertainty for financial planning and rarely quantifies it, so we applied conformal prediction to the problem. The work reviews the underlying theory — calibration, quantile regression, and threshold-calibrated guarantees — then builds a framework that pairs it with standard ML models to produce statistically rigorous prediction intervals. It holds the target coverage rate across six demographic subgroups defined by gender and region while adapting interval width to local uncertainty, which makes fairness a coverage constraint rather than an afterthought.

      Ansh Nagwekar, Arush Mehrotra CIS 6250: Theory of Machine Learning

  3. Open Core Ventures

    2022 – 2025

    Investment Research

    Technical due diligence and market research across 250+ open-source projects spanning DevOps, gaming, and GenAI. Advised on seed investments for various startups including Kernelize, MCPJam, Gobii, SurfSense, Lucenia, and Phaser — see the full portfolio.

  4. Newristics

    2020 – 2021

    Machine Learning Research

    Trained SVMs to 80%+ accuracy identifying 40+ cognitive biases in pharmaceutical marketing material, and wrote a research report on NLG style transfer with large language models to guide automation of AiGILE.

Back

Quotes

  1. τὰ ἐφ' ἡμῖν, τὰ οὐκ ἐφ' ἡμῖν

    (ta eph' hay-MEEN, ta ook eph' hay-MEEN)

    “What is up to us, and what is not up to us.”

    Epictetus, Enchiridion

  2. “I'm nice at ping pong.”

    Kanye West

  3. “Do you want to take a leap of faith, or become an old man, filled with regret, waiting to die alone?”

    Saito, Inception Christopher Nolan

  4. “Of everyone on Golden State, open shot, the fate of the universe on the line, the Martians have the death beam pointed at earth, you better hit it — I want Iguodala.”

    Max Kellerman, First Take

  5. “Above all, do not lie to yourself. The man who lies to himself and listens to his own lie comes to such a pass that he cannot distinguish the truth within him, or around him…”

    Father Zosima, The Brothers Karamazov Dostoevsky

  6. “If I told you that a flower bloomed in a dark room, would you trust it?”

    Kendrick Lamar, Poetic Justice

  7. “You know what kind of plan never fails? No plan. No plan at all. You know why? Because life can't be planned.”

    Kim Ki-taek, Parasite Bong Joon-ho

  8. “Sometimes I'll start a sentence and I don't even know where it's going. I just hope I find it along the way.”

    Michael Scott, The Office

  9. “This city is afraid of me. I have seen its true face.”

    Rorschach, Watchmen Alan Moore

  10. “My child-like creativity, purity and honesty is honestly being crowded by these grown thoughts.”

    Kanye West, POWER

  11. “Murph, don't let me leave.”

    Cooper, Interstellar Christopher Nolan

  12. “Tight, tight, tight! Yeah!”

    Tuco Salamanca, Breaking Bad

  13. “Lost time is never found again.”

    Benjamin Franklin, Poor Richard's Almanack

  14. “Reality is far more vicious than Russian roulette. First, it delivers the fatal bullet rather infrequently, like a revolver that would have hundreds, even thousands of chambers instead of six. After a few dozen tries, one forgets about the existence of a bullet, under a numbing false sense of security. Second, unlike a well-defined precise game like Russian roulette, where the risks are visible to anyone capable of multiplying and dividing by six, one does not observe the barrel of reality. One is capable of unwittingly playing Russian roulette — and calling it by some alternative ‘low risk’ game.”

    Nassim Nicholas Taleb, Fooled by Randomness

  15. “Suddenly I realised that I was no longer driving the car consciously. I was driving it by a kind of instinct, only I was in a different dimension.”

    Ayrton Senna, 1988 Monaco Grand Prix

  16. “No half-hearted oinks. I want full-hearted oinks.”

    Logan Roy, Succession

  17. “What comes next will be marvelous.”

    The Blessed Madonna, Marea (We've Lost Dancing) Fred again..

  18. “Skadoosh.”

    Po, Kung Fu Panda

  19. “I'll just keep playing back
    These fragments of time
    Everywhere I go
    These moments will shine”

    Daft Punk, Fragments of Time

  20. “I saw the best minds of my generation destroyed by madness…”

    Allen Ginsberg, Howl

  21. “O Captain! My Captain!”

    John Keating, Dead Poets Society Peter Weir

  22. “Here's to the ones who dream
    Foolish as they may seem
    Here's to the hearts that ache
    Here's to the mess we make”

    Mia, La La Land Damien Chazelle

  23. “Brazil, Morocco, London to Ibiza
    Straight to L.A., New York, Vegas to Africa
    Dance the night away
    Live your life, and stay young on the floor”

    Jennifer Lopez, On the Floor Pitbull

Back