Hi, I'm Aarav. I'm a computational neuroscience and machine learning researcher interested in how biological neural systems generate intelligent behavior, and how those principles can inform new AI systems.
I'm currently a computational neuroscientist at Eon Systems PBC working on embodied Drosophila brain models, with our goal being to one day replicate human consciousness, and a researcher at Johns Hopkins University in the Dynamical Intelligence Group studying the computational theory of predictive grid cells in the medial entorhinal cortex (MEC). Previously, I was a Member of Technical Staff at Humanis, where I worked on building probabilistic world models of complex physical systems using neurosymbolic AI, and a summer researcher at Harvard University's Kempner Institute training RNN agents for odor plume tracking.
I'm a junior at Tompkins High School in Katy, Texas, where I'm ranked 5th out of 800+ students with a 4.0 GPA. I'm also a USACO Gold competitor, a Science Olympiad state medalist, and the founder of my school's AI Club and Engineering Club.
MapShift was accepted to the NeurIPS 2026 Evaluations and Datasets Track as a poster! I'll also present my TMLR paper through the Machine Learning Reproducibility Challenge (MLRC) at NeurIPS. Read the blog post →
My paper Auditing Closed-Loop Learning in Recurrent Neural Networks was published in Transactions on Machine Learning Research (TMLR).
Eon's embodied Drosophila brain emulation — the first multi-behavior embodied brain model — reached 100M+ views across X, LinkedIn, YouTube, and more, and drew a reaction from Elon Musk. Read the story →
I'm passionate about computational neuroscience, connectomics, embodied neural models, and deep reinforcement learning. My work focuses on understanding how biological neural circuits give rise to behavior, using both data-driven simulations and biologically inspired AI agents.
NeurIPS Poster (2026)
An executable benchmark for controlled post-intervention evaluation of embodied world models. Matched environment pairs across metric, topology, dynamics, and semantic interventions separate stale map reuse, belief update, and post-change planning, revealing failures and ranking changes hidden by same-environment evaluation.
Under review at PNAS
Oral presentation at the NeurIPS 2026 Workshop on Symmetry and Geometry in Neural Representations.
Predictive grid coding appears in rodent neural recordings and RNNs trained only for path integration, without an explicit predictive objective. Our results suggest that predictive grid cells may help maintain and update grid cell population dynamics, rather than necessarily supporting explicit planning.
Transactions on Machine Learning Research (TMLR), 2026. Accepted for Presentation at MLRC, NeurIPS 2026.
A claim-level reproducibility audit of closed-loop RNN learning across 675 paired runs. Transient deployed-loss peaks and coupled-system stability signatures reproduce robustly, while final-loss gaps and transfer across architectures and tasks are more conditional.
NeurIPS AI for Science Workshop, 2025
We use deep RL to train biologically inspired RNN agents to navigate to odor sources, comparing walking and flying modes. Walking agents develop fine-scale orientation strategies and compact neural representations, while flying agents use broad sweeping turns with higher-dimensional dynamics.
Society for Neuroscience (SfN), 2025 — Poster
A connectome-constrained Drosophila brain model that accurately predicts neural activity and controls behavior in a virtual environment. I contributed to the brain embodiment component of the project at Eon Systems, whose embodied fly emulation later reached 100M+ views online.
Off the screen, I play the violin.
A dedicated watcher — Death Note is a forever favorite.
And I make a serious bowl of matcha.
Academic CV
A complete record of my research, experience, publications, awards, and academic activities.
Mailing List
Get an email when I publish a new blog post or share a major project update.
