Stone Tao

Highlights

3x MIT Battlecode Finalist, best result: 1st out of solo competitors, 5th overall out of 600+ graduate to HS students globally.
Launched the Lux AI Challenge with Kaggle, with 800+ teams and github stars, 9000+ submissions after 2 months.
Placed top amongst students in graduate class competitions in deep learning, computer vision, and reinforcement learning.
Self-driven, persistent, full stack software engineer, AI and HCI researcher.

Developed / researched high precision and accurate white-box function approximation using deep neural nets, decision trees, and boosting. Helps automatically migrate slow, complex, hand-built calculators in old systems to new systems, and learn fast differentiable approximations of these functions for analysis purposes.
Results beat LGBM, deep neural nets, and other methods by 100x or by being feasible in high dimensions.
Developed OCR+NLP tools for analysis of insurance documents for automatic categorization of insurance types and their properties
Used Pytorch, Jax, Flax, Optax, scikit learn, scipy, Pandas, GCP

Researching Reinforcement Learning (RL) and Robotics under Professor Hao Su. Currently researching skill translation between different morphologies. Previously built gym environments and systems for the ManiSkill challenge to benchmark RL, CV, and robotics on SAPIEN, a simulated part-based interactive, 3D environment. Work accepted to NeurIPS 2021 https://arxiv.org/abs/2107.14483

Worked full stack on feature workflows, semantic patches, and conflict handling to enable state independent scheduling of feature flagging, allowing users to release complex features with confidence. Developed a REST API to enable an approval review system for feature flagging, a feature requested by LaunchDarkly's largest business customers. Used Go, React, and Typescript.

Researching at the intersection of AI and HCI. Currently researching NLP summarization methods, tree learning models, and how to introduce crowdwork to improve models, fairness, accountability, and transparency.

Reinforcement Learning Gym and Library in Typescript - Apr. 2021: github.com/StoneT2000/rl-ts

Implements a gym interface and algorithms like PPO and DQN in Typescript for reinforcement learning on browsers and Node.js.

Dimensions - Generalized AI Competition Framework – Apr. 2020: github.com/StoneT2000/dimensions

Allows users to easily create language-agnostic AI competitions. Provides Google Cloud and MongoDB integrations to scale up a competition in 3 lines of code.
Being used in a collaborative effort with Kaggle to run an AI competition called the Lux AI Challenge: https://lux-ai.org/

MIT Battlecode (AI Competition) Finalist: Made finals 3 times in a row (2019-2021), competing against over 600 teams of high school to graduate students, won the Five Rings adaptive strategy award for spearheading an influential strategy in 2021.
Graduate Robotics and RL Course. 2nd out of 20+ graduate students in RL competition using PPO and Random Network Distillation.
Graduate Computer Vision (ML Meets Geometry). Highest placing undergraduate student on 3D segmentation and pose estimation tasks, using PointNet / PointNet++, Frustum PointNet, Faster-RCNN etc.

Undergraduate: University of California San Diego, Graduation Date: Jun. 2023

B.S. Computer Science, Cognitive Science (double major); Math minor (intended); GPA: 3.98/4.0 - Provost Honors List
Graduate Courses: Computer Vision (ML meets Geometry), Robotics and RL, Differentiable Programming, Recommender Systems
Undegraduate Courses: Honors Linear Algebra & Honors Calculus sequence, Decision Making in the Brain, Advanced Data Structures, Advanced Optimization Methods for Data Science, Design and Analysis of Algos, Data Science in Practice, Computer Architecture: Software Perspective, Operating Systems
Activities: Founding president of ACM AI at UCSD; Member of TBP Honors Engineering Society at UCSD

Programming Languages: Typescript, Python, SQL, Go, C, C++, Java, PHP, Javascript
AI: RL, 3D CV (Object Detection, Segmentation, Pose Estimation), Deep Learning, Tree Learners, Boosted Trees
Frameworks/Engines/Libraries: Pytorch, Tensorflow, Pandas, scikit-learn, Jax, Flax, OpenAI Gym, RLLib, Numpy, Matplotlib, Seaborn, Node.js, React, MongoDB, Express.js
Tools: Docker, Google Cloud, Jupyter Notebook, Git, Adobe Photoshop