Projects

Things I've built

Two systems I built outside of work — one in systematic trading infrastructure, one in multi-agent orchestration.

Oct 2024 — PresentSystematic Trading Research Platform

A modular Python framework for quantitative portfolio management， the full stack from factor estimation to constrained optimization to walk-forward execution, with every step grounded in theory and verified by tests.

Factor Models
PCA/PPCA-based statistical factor models with missing data handling (listwise delete, forward fill, EM), loading imputation, real-time diagnostics tracking.
Risk Models
Fundamental risk models, currency risk rebasing, linked multi-horizon covariance estimation.
Volatility
GARCH(1,1), DCC (Dynamic Conditional Correlation), realized volatility estimators, stochastic time-varying uncertainty (STVU).
Portfolio Optimization
Mean-Variance Optimization (three equivalent formulations), Factor-Mimicking Portfolios, factor-neutral idiosyncratic-space portfolios, IC / Fundamental Law analytics.
Constrained & Robust MVO
CVXPY solver supporting 10 constraint types (long-only, turnover, tracking error, factor exposure, etc.) with progressive degradation and KKT dual diagnostics. 5 robust approaches for estimation error.
Execution
Walk-forward backtesting with transaction cost modeling, Sharpe CI, autocorrelation adjustment, backtest hygiene checks.

The constrained MVO solves the standard mean-variance problem with active weight penalties and a full constraint set:

$$\max_{w} \; \alpha^T w \;-\; \frac{\lambda}{2}\, w^T \Sigma\, w \quad \text{s.t.} \;\; Aw \leq b,\;\; \|w - w_0\|_1 \leq \tau$$
where \(\alpha\) is the expected return vector, \(\Sigma\) the risk model covariance, \(\lambda\) the risk aversion parameter, and \(\tau\) the turnover budget. The solver uses a sparse epigraph formulation for \(n=2000+\) assets and falls back through a Cholesky → ridge → pseudo-inverse chain for numerical stability.

For spiked covariance structures (low-rank factor model + diagonal idiosyncratic), the platform uses the Woodbury-Sherman-Morrison identity for O(n) solves instead of inverting the full \(n \times n\) matrix:

$$(\Sigma_{\text{idio}} + B F B^T)^{-1} = \Sigma_{\text{idio}}^{-1} - \Sigma_{\text{idio}}^{-1} B \left(F^{-1} + B^T \Sigma_{\text{idio}}^{-1} B\right)^{-1} B^T \Sigma_{\text{idio}}^{-1}$$
This reduces the inversion from \(O(n^3)\) to \(O(nk^2)\) where \(k \ll n\) is the number of factors — critical when optimizing over 2000+ assets with daily rebalancing.
28K
Lines of code
244
Test assertions
MC
Verified
Python CVXPY NumPy XGBoost LSTM Transformer Walk-Forward Monte Carlo

2024 — PresentMulti-Agent Inference Framework

A LangGraph-orchestrated multi-agent system where specialized agents each handle a domain, coordinate through belief propagation over a DAG, and produce a unified assessment with explainable reasoning. The application was clinical risk detection, but the architecture is domain-agnostic — it's the same routing-and-fusion pattern applicable to any multi-signal decision problem.

Knowledge distillation + mixed-precision quantization compressed the system 93% in size while keeping accuracy within 1.2% of the original. An adaptive router directs easy cases to lightweight models and escalates high-entropy cases to larger ones. Each agent incorporates RAG for evidence-grounded outputs. Validated on N=2,847 across 3 sites, deployed on Kubernetes with MCP-based integration, processing 500+ assessments daily at 99.7% uptime.

0.963
AUC-ROC
73%
Cost reduction
<2s
Latency
LangGraph Multi-Agent Knowledge Distillation RAG MCP Kubernetes
Background

Education & toolkit

Stanford University
B.S. Mathematics
2017 — 2022
Stochastic Calculus · Machine Learning · Convex Optimization · Mathematical Finance · Numerical Linear Algebra · Algorithms · Distributed Systems
🥇 WorldQuant Challenge — Gold
🤖 Mathematical Robot Competition — 1st
Languages
Python · C++17 · R · MATLAB · SQL
ML / AI
PyTorch · TensorFlow · XGBoost · LightGBM · LangGraph · HuggingFace · QLoRA · vLLM
Quant
NumPy · pandas · statsmodels · cvxpy · zipline · Bloomberg · IBKR
Infrastructure
Docker · Kubernetes · Redis · PostgreSQL · Git · Linux · gRPC