Experience

Cisco's treasury team manages the company's cash, investments, and share buybacks — a portfolio whose risk gets measured every day. Over the summer I built and upgraded the team's risk and analytics tooling, from the Excel models the desk runs on to a Python (pandas, NumPy) command-line tool the whole team uses:

• Value-at-Risk model (Excel) — upgraded the team's risk forecasting to incorporate Monte Carlo simulation and implied volatility, so the estimates reflect what the market expects rather than what it did last quarter.
• Counterparty bank analytics (Excel) — evaluates roughly twenty counterparty banks on our trades and their long-run markups.
• Transaction classifier — K-Means clustering distilled into a decision-tree surrogate, so every classification stays explainable. It categorized 100,000+ transactions and strengthened the accuracy of the team's transaction data, with its recommendations reviewed and adopted by the cash team. A risk number is only as good as the data underneath it.
• Agentic AI workflow — turns a static buyback sheet into full broker analysis automatically.

I design and maintain purduecrew.com, the website for Purdue's rowing team, and I'm building a portal where athletes, coaches, and parents can interact with a database of athlete metrics, race schedules, and workout plans.

RiskDesk runs the cycle a bank's market-risk function runs every night, for a mock three-desk trading book — cash equities, FX spot, and US Treasuries. Ingest market data, gate it on quality, revalue the book, compute VaR and Expected Shortfall, check limits, replay a catalog of crises, score yesterday's forecast against what actually happened, and publish it all to a dashboard. The book is invented; the market data is real and public, back to 2007.

• Hand-rolled math — historical-simulation and EWMA-filtered VaR, 97.5% Expected Shortfall with stressed calibration, closed-form bond pricing with key-rate DV01s off a bootstrapped zero curve, Black-Scholes and Greeks, GARCH(1,1) by QMLE, and the Kupiec and Christoffersen likelihood ratios. Only optimizers and distribution functions are imported — no QuantLib, no statsmodels.
• Validated, not asserted — 750 out-of-sample days of backtesting, 6 exceptions against 7.5 expected (Kupiec p = 0.57), GREEN on the Basel traffic light for both methods, plus a real P&L-attribution test.
• Governed like a model, not a script — every tunable frozen in one config, each run stamped with its git SHA, a champion/challenger cycle against pre-registered promotion criteria, and an SR 11-7-structured model document whose limitations are measured rather than just listed.
• Built to run again tomorrow — idempotent batch runs under an advisory lock, a data-quality gate that downgrades a run rather than failing quietly, and a committed market snapshot so CI reproduces the firm's VaR to the cent on every push.

I wrote up what it taught me — mostly about the pipeline rather than the math — in what breaks when a risk model has to run every night. Code on GitHub. Educational demonstration on public data; not investment advice.