I study the geometric structure of “intelligence”. Currently my work is on AI interpretability, specifically studying the geometry of how models internally represent a task.
I am an MSCS student at UIUC working on computational mechanics and decision theory, to better understand how a model represents and builds its understanding of the world. I am deeply inspired by the work on Why Nations Fail, and my current hypothesis focuses on seeing how AI models can better capture the structures of nations and institutions.
Research
My current research, begun at Simplex AI Safety, is on how LLMs form and update representations of latent belief states while making predictions. My goal at school is to show that models can learn and perform inference over structures identified by economics, game theory, political science, and complex-systems research.
- Large Language Models Develop Belief State Geometry In-Context NeurIPS 2026 (Poster) · *co-first authors · paper
- How Reliable are Causal Probing Interventions? IJCNLP-AACL 2025 · earlier version at NeurIPS 2024 IAI Workshop · paper
- A Self-Supervised Miniature One-Shot Texture Segmentation (MOSTS) Model for Real-Time Indoor Drivable Area Segmentation IEEE CASE 2025 · paper
- A Convolutional Neural Network Based Vision-Proprioception Fusion Method for Robust UGV Terrain Classification IEEE Robotics and Automation Letters · presented at IEEE/RSJ IROS 2021 · paper
- Comparison of Dynamic and Kinematic Model Driven Extended Kalman Filters (EKF) for the Localization of Autonomous Underwater Vehicles arXiv preprint · paper
Experience
-
Jan — Jul 2026
Research Fellow (SPAR), Simplex AI Safety Led a project on in-context learning. Showed LLMs reconstruct their Bayesian belief state in-context, recoverable by linear probes (R² > 0.998) and causally used for prediction.
-
2023 — 2026
Embedded Software Engineer, Qualcomm Wrote Linux kernel software for hardware-accelerated packet processing on 5G modems. Built a crypto hardware-offload path (~10% TCP throughput) and cut buffer memory by over 80%.
-
2021 — 2023
Undergraduate Researcher, IBM-Illinois Centre Causal interventions on Transformer representations across GPT-2, Pythia, and Llama, showing common methods overestimate causal significance. Built BM25 plus Transformer semantic search for CS 410.
-
2020 — 2021
Software Engineering Intern, Caterpillar Built a machine-utilization forecasting system that guided executive decisions through COVID-19, winning the 2020 CIO Award. Led the Snowflake migration, cutting runtime >99% and cost ~66%.
-
2019 — 2023
Undergraduate Researcher, AUVSL Lab, UIUC Self-supervised texture segmentation with Perlin-noise collage masking (85.7 mIoU at 93 FPS), and vision-proprioception fusion for robot terrain identification (96.4% accuracy).
M.S. Computer Science (thesis track), University of Illinois at Urbana-Champaign, expected 2027 · Teaching Assistant, CS 411: Database Systems.
B.S. Computer Engineering, University of Illinois at Urbana-Champaign, 2023 · James Scholar, Dean's List.
Projects
- Belief State Geometry → Interactive visualization of belief states in hidden Markov models like Mess3, and how their geometry changes as you sweep the model's parameters.
- MOMENT Event database and app for multiple universities, serving 1,200+ users and integrated with the official UIUC app. Built with ROKMETRO under Dr. Sanjay Patel.
- Early Wildfire Detection using 5G Massive IoT 3rd place at QHACK (Qualcomm) — two-tiered edge server architecture on AWS enabling high-frequency data collection from millions of sensors while minimizing storage costs.
- AIR Best financial hack at HackIllinois, presented by Capital One — tracks fingers to create and sign usable checks in the air, using Google Cloud Vision and Checkbook.io.
Geopolitics
An open-source intelligence project tracking global military operations and great-power dynamics. Choose a view: