Building intelligent software that transforms data into decisions.

A predictive model for the 2026 FIFA World Cup, extending an earlier passing-network analysis into a full forecasting pipeline trained on five international tournaments.

Problem

The 2026 field expands to 48 teams, and many new entrants from the AFC, OFC, and CONCACAF have no prior World Cup event data — a model trained only on past World Cups has no way to cover most of the field.

Solution

Each team's features — progressive passing, press intensity, expected-goals differential, defensive activity — are pulled from its best-available tournament (World Cup, Euros, Copa América, or AFCON) and converted to percentile ranks so scores compare fairly across competitions of different strength. Teams with no coverage at all get a conservative confederation-average estimate instead of a guess. The logistic regression model is validated with leave-one-tournament-out testing, then calibrated and rescaled so all 48 teams' semifinal probabilities sum to exactly 4.

Outcome

Consistent out-of-sample AUC above 0.70 across held-out tournaments, with expected-goals differential emerging as the strongest single predictor — offensive playmaking alone doesn't win without defensive quality behind it.

A win-probability model for Virginia Tech football, paired with four Claude agents that explain, contextualize, and critique the model's own predictions rather than just outputting a number.

Problem

A win-probability percentage alone doesn't tell you why — most sports prediction tools either bury you in raw stats or skip the reasoning entirely, and models rarely check their own blind spots.

Solution

A logistic regression model trained on 10 seasons of FBS games (SP+ ratings, Elo, roster talent, rest days) produces the base win probability. Four Claude agents then reason over that output independently: one explains it in scout language tied to specific units, one assesses transfer-portal and recruiting context, one compares it against sportsbook lines, and a fourth runs deterministic checks — including a grounding scan that flags any number an agent cites but wasn't actually given — that can only lower the final confidence tier, never raise it.

Outcome

65.4% held-out accuracy (AUC 0.70) against a 58.9% baseline, shipped as a live dashboard defaulting to VT's next real matchup. Along the way, a coaching-change feature that looked like it should matter tested out with a coefficient indistinguishable from zero — and was left out rather than kept for a good story.

A live market dashboard anchored on the S&P 500, built to explain why the market moved instead of just displaying numbers, with a stock watchlist covering fundamentals and forward estimates.

Problem

A grid of live tickers doesn't tell you anything on its own — a move only means something in the context of whether the usual relationships between assets are holding or breaking down that day.

Solution

Each move in the S&P is checked against the VIX, the 10-year yield, and the dollar using a trailing 90-day rolling correlation, flagging whether today's reaction confirms or diverges from the historical relationship. A custom sentiment score (built from VIX percentile, price trend, and gold's relative strength, not a scraped black box) feeds a sector-by-sector signal table. A watchlist tracks free cash flow, CapEx, profit margins, and operating income over time per stock, plus forward analyst EPS estimates.

Outcome

Caught a real bug during testing where two of the three correlation factors were silently returning null because their price feeds report timestamps in different exchange timezones — fixed by normalizing to trading day before joining.

Reconciliation automation built during my internship at Investcorp, comparing financial transaction records between the fund accounting system and bank statements.

Problem

Reconciling records between two independent systems by hand is slow, and mismatches are easy to miss until they compound.

Solution

A Python pipeline using Pandas and NumPy matches transaction records and flags mismatches automatically, with validation logic and structured reporting output to support financial review.

Outcome

Reduced manual comparison work by surfacing mismatches automatically instead of requiring a full line-by-line review.

An interactive dashboard tracking technology spending across business units, built to replace a static monthly reporting cycle.

Problem

Leadership needed a current view of technology spend across business units without waiting on the next manual report.

Solution

Built interactive dashboards using DAX measures, slicers, and drill-through functionality, connected directly to live financial databases so figures reflect same-day activity rather than a monthly export.

Outcome

Replaced static monthly decks with a self-serve tool finance and line-of-business leadership could query directly in planning meetings.