Mixture-of-Steering Vectors
Prompt-conditioned activation steering for mitigating factual hallucinations in language models.
About
I'm a CS student at Stanford working across mechanistic interpretability, multi-agent systems, and memory in AI systems.
I'm currently at Salesforce Research, working on pattern memory for coding agents, after previously building a multi-agent workflow optimizer on the same team.
I've also worked on steering vector methods for LLM interpretability (MoSV, presented at ICML 2026) and satellite collision avoidance at SAIL. I also contract with state departments of transportation, training models and building data pipelines for wildlife conservation and transportation safety.
I'm interested in how models and agents build, use, and steer internal representations over time and what that means for understanding and controlling their behavior as they scale. That question shows up in different forms across what I work on, from steering vectors in a single model to how memory and coordination emerge in multi-agent systems. I'm actively looking to go deeper on this through more focused research, and I'm always happy to talk to anyone thinking about similar questions.
Experience
Salesforce AI Research, AI Software Engineer
2025 to presentWorking on memory for coding agents and previously built optimization frameworks for multi-agent workflows.
Stanford Intelligent Systems Laboratory, Student Researcher
2025 to presentPreviously developed optimization methods for ground-station placement and communication across large satellite constellations. Currently working on satellite collision avoidance.
Iyarkai, Co-founder
2021 to presentBuilding computer-vision systems for wildlife monitoring, including thermal-imaging deployments with the Washington State Department of Transportation.
UC Davis Road Ecology Center, Applied Science Researcher
2021 to 2025Developed and deployed RADIS, an edge-AI system for detecting wildlife near highways and warning drivers.
UC Davis School of Veterinary Medicine, Engineering Intern
2022Designed embedded, safe deterrent systems for wildlife conservation applications.
Research & Projects
Optimization for Large LEO Constellations
Optimization methods for ground-station placement, communication coverage, and safe operation of large low-Earth-orbit satellite constellations.
Automated Thermal Wildlife Monitoring
Synthetic-data and computer-vision pipeline deployed for wildlife monitoring at Washington highway crossings.
RADIS Wildlife Detection System
An edge-computing animal detection and driver-warning system deployed alongside highways in Nevada.
Biomimetic Drag Reduction for Heavy Vehicles
Computational and wind-tunnel study of boxfish-inspired trailer geometry, published in Intersect.
Other projects
Dynamic Portfolio Optimization with PPO
Reinforcement learning for daily asset allocation under transaction costs and changing market conditions.
Phase Recognition in Surgical Videos
A lightweight causal Transformer and multi-video finetuning strategy for long-form surgical video understanding.
NBA Player Statistics Predictor
Singular-value-decomposition models for predicting player performance from historical statistics.
Ora Animal Deterrent System
A configurable audiovisual warning system for wildlife safety, later adapted to help protect collection animals at the San Diego Zoo Safari Park from mountain lions.