Ishaan Gupta

cs + ai @ stanford, research at stanford neuroai / nasa genelab

github linkedin email

Experience

Stanford NeuroAI Lab

Research Intern, World Models · Stanford, CA · May 2026 – Present

▪First-authoring a mechanistic interpretability study of PSI, a 131K-unit multimodal world model; isolated 24 causally necessary units across five modalities. Presented at Stanford CURIS.
▪Built sparse autoencoders and block-sparse featurizers in PyTorch with distributed multi-GPU training, enabling per-modality activation steering.

NASA GeneLab

SWE Intern, AI/ML Analysis Working Group · Mountain View, CA · May 2024 – Present

▪Built a deep learning classification pipeline over 15K+ Arabidopsis gene-expression samples, improving accuracy 23% over baseline.
▪Deployed an automated preprocessing framework on Airflow, AWS Lambda, S3, and Docker, cutting manual intervention time 60%. Publishing in Scientific Reports, 3rd author.

UC Berkeley AQMEL

Software + Research Intern · Berkeley, CA · June 2023 – June 2025

▪Owned the software stack for an EPA-funded air-quality sensor network, from IoT telemetry ingestion to wildfire smoke exposure forecasting.
▪Shipped an ML regression pipeline (XGBoost, TabPFN) forecasting PM2.5 up to 72 hours out, 8.1% more accurate than legacy models.

Projects

Shark Health Modeling

Computer vision + RL for white and leopard sharks, Monterey Bay · Featured on CNN International — Tech4Good

CNN International — Tech4Good →

Great white shark filmed head-on underwater in Monterey Bay

Research and CNN film crew on a boat during shark tagging fieldwork in Monterey Bay

Developed neural networks with YOLO detection and pose segmentation to track white sharks in Monterey Bay from drone and dive footage.

Now using reinforcement learning with PPO to identify and label shark scars and injuries in leopard sharks.

WSN, Nov 2025NEPSS, Mar 2026

Projects, Cont.

3D World Models for CAD

3D world models for editable, manufacturable geometry

Building a generative system that turns natural-language and reference-image prompts into editable, manufacturable CAD geometry.

We've begun development on postraining existing frontier models to see how well we perform on CadGenBench. Current work spans the geometry representation, a differentiable constraint solver for edits, and an evaluation suite that scores outputs on printability.

Backed by the 1517 FundMore coming soon.