Professional Experience

Lead the applied research vertical of a team building an agentic AI platform, and build the production retrieval and evaluation systems it runs on.

Responsibilities

• Lead the team’s applied research vertical, with a research agenda centered on agentic orchestration and execution-grounded evaluation
• Prototyping agentic execution, in which models draft and execute plans, use structured errors to revise them, and recover failed runs by selecting alternative plans
• Redesigned and shipped the platform’s document retrieval system, combining filename-aware routing with content retrieval across document families and versions
• Built the evaluation layer for the platform’s LLM-generated SQL, measuring the correctness, stability, and groundedness of both queries and their executed results
• Re-engineered the group’s evaluation suite for concurrent execution, cutting a full run from about 3 hours to about 8 minutes
• Built an internal tool that provisions a vLLM inference endpoint for any model in about 30 seconds, now used across the group

Technologies & Skills

• Agentic systems, retrieval-augmented generation, and LLM evaluation
• Python, LangGraph, and vLLM

Researched machine learning methods for particle physics with Professor Benjamin Nachman, focused on unfolding: statistically recovering the true distributions of physical quantities from noisy, high-dimensional, detector-distorted collider data.

Responsibilities

• Developed generative and adversarial models to reconstruct latent distributions of jet observables from high-dimensional, noisy collider data
• Invented machine learning algorithms for signal recovery and inference under uncertainty, including Moment Unfolding, Reweighting Adversarial Networks, and Neural Posterior Unfolding
• Implemented deconvolution code enabling statistical estimation in distorted measurement spaces, reducing run-time cost by ~100x while maintaining precision
• Collaborated with the CMS experiment on data analysis at the Large Hadron Collider
• Mentored undergraduate researchers on machine learning projects
• Published in Physical Review X, Physical Review D, the European Physical Journal C, the Journal of Instrumentation, the Annals of Applied Statistics, and NeurIPS, and presented at CERN, KIAS, NeurIPS, and APS meetings

Technologies & Skills

• Python, PyTorch, TensorFlow, and JAX
• Generative models, normalizing flows, and adversarial networks
• Statistical inference, deconvolution, and uncertainty quantification
• High-performance computing on NERSC systems

Developed and fine-tuned language and vision-language models for client-specific applications at Emissary AI.

Responsibilities

• Developed and fine-tuned language and vision-language models for client-specific applications, including code completion and image-based IP violation detection
• Implemented a GRPO-based reinforcement learning pipeline for fine-tuning language and vision models on code completion and image-based IP violation detection
• Improved model accuracy from around 30% to over 90% for redacted medical message classification through architecture optimization and training refinements
• Refactored the training codebase for distributed multi-GPU systems, improving scalability and throughput for fine-tuning transformer models

Technologies & Skills

• Large language models and vision-language models
• Model fine-tuning, reinforcement learning (GRPO), and architecture optimization
• Distributed multi-GPU training
• Python, PyTorch, and deep learning frameworks

Served as Associate Instructor in the UC Berkeley Physics Department, teaching a range of undergraduate courses.

Courses Taught

• Physics 89: Introduction to Mathematical Physics
• Physics 88: Data Science Applications in Physics
• Physics 77: Introduction to Computational Techniques in Physics
• Physics 8B: Introductory Physics (Head Associate Instructor)
• Physics 7B: Physics for Scientists and Engineers

For my full teaching record, see the teaching page.

Technologies & Skills

• Curriculum development and course instruction
• Python, data science, and mathematical and computational physics pedagogy

Worked on systematic trading research and execution at Bridgewater Associates, developing probabilistic models and strategic frameworks to optimize investment decisions.

Responsibilities

• Developed Bayesian hierarchical models to predict liquidity for trade execution, enabling data-driven selection over trader decisions and reducing transaction costs
• Implemented probabilistic frameworks for systematic trading strategies to minimize transaction costs and alpha leakage subject to risk constraints, under varying market conditions
• Backtested and optimized execution strategies under uncertainty, forecasting transaction costs on sparse data using partial pooling

Technologies & Skills

• Bayesian hierarchical modeling and partial pooling
• Quantitative trading strategy design and execution
• Python, R, and statistical computing
• Financial modeling, liquidity analysis, and portfolio optimization

Conducted theoretical and computational research at Microsoft Research, advancing mathematical foundations of non-local field theory and its applications to optimization.

Responsibilities

• Collaborated with Jaron Lanier (Chief Unifying Scientist) on developing non-local field theory from matrix models
• Bridged discrete and continuous structures through new techniques applicable to optimization problems
• Applied stochastic calculus and operator theory to establish quantitative relationships between local and non-local dynamics
• Designed and executed numerical simulations to validate predictions about high-dimensional operators
• Quantitative modeling and analysis incorporated and acknowledged in Lanier et al. (2022)

Technologies & Skills

• Stochastic calculus, operator theory, and matrix models
• Theoretical physics and applied mathematics
• Numerical simulations and computational modeling
• Python, MATLAB, and high-dimensional analysis

Contributed to the development of advanced eye-tracking software at Purple Gaze Inc., building AI-driven algorithms and real-time image processing pipelines for production deployment.

Responsibilities

• Engineered AI-driven glint detection algorithms for eye-tracking software, significantly improving detection accuracy
• Wrote production-quality code in Python and C for real-time image processing applications
• Collaborated within an agile startup environment, delivering features under tight deadlines while ensuring high code quality

Technologies & Skills

• Python, C, and real-time image processing
• Computer vision and AI-driven algorithm design
• Agile software development practices
• Performance optimization and production deployment