Andrei Piterbarg
CS + ML · University of Pennsylvania
I do research at the GRASP Lab building adversarial agent systems, and I am collaborating with Prof. Damek Davis and Prof. Terence Tao on open problems in harmonic analysis and optimization.
I am also currently working with Alexander Sokol, studying autoencoder yield curve manifolds as a follow-up to Lyashenko, Mercurio & Sokol (2024).
I am interested in building Machine Learning systems grounded in research, especially in understanding how research and applied ML can work together.
Designing experiments on multi-agent information aggregation, evaluating convergence and regret across synthetic and structured prediction tasks.
Developed a Numba-JIT and MOSEK framework for bounding autoconvolution constants, combining Lasserre SDP hierarchy relaxations with simplex cuts.
Studing the structure of low-dimensional latent manifolds capturing term-structure dynamics across markets and currencies.
Built Graph-RAG system for customer onboarding with hallucination mitigation and optimal indexing. Implemented retrieval pipelines and evaluation harnesses, improving performance via prompt and index tuning.
Built transformer with prioritized experience replay that won 1st place in the Quantitative Analytics track and 3rd place overall at the Penn × Anthropic Hackathon. The model weighs electricity cost, network connectivity, climate, and satellite/environmental data (Planet Labs API).
The system correctly identifies regions like Northern Virginia, Oregon, and Iceland while rejecting high-cost or disaster-prone areas. Built with PyTorch, scikit-learn, and FastAPI.
- Penn × Anthropic Hackathon: 1st place in Quantitative Analytics, 3rd place overall
- UK National Coding Cipher Challenge: Individual Winner
- UK Math, Informatics and Physics Olympiads: Gold Award (4×)