Sept 2026 – Present · Personal Project

• Deployed end to end on Next.js, Cloudflare Workers and R2 at calibratedsports.com, with a 552,646 row strategy backtest engine, an append only forecast ledger graded in public, and a contract gated export covering 1,270 tests
• Built a platform converting prediction market ladders and sportsbook odds into full outcome distributions for every priced player, fitting isotonic survival curves and a Gaussian copula Monte Carlo over 3,988 players and 7,309 games from 1999 forward
• Pre registered and published 18 hypotheses under walk forward validation, measuring a 2.43 pp over side bias (95% CI 1.72 to 3.13) across 44,198 settled props and scoring the in house model against the closing line on 14,857 out of sample predictions

Personal photography portfolio on an interactive 3D globe. EXIF-driven metadata, filterable gallery, journey playback across 200+ photos and 22 places.

• Replicated the S&P 500 with a 30–50 stock subset: extracted the top 5 PCA factors from a 488-stock universe, then selected leader stocks by factor correlation and residual regression until R² ≥ 95%, following Jiang & Perez (2021)
• Built max-Sharpe portfolios with beta-proxied expected returns and ±25% weight bounds under three constraint sets (unconstrained, no short-selling, shrunk beta); the no-short portfolio tracked best, consistent with non-negativity acting as regularization (Jagannathan & Ma, 2003)
• Stress-tested across the 2022–23 regime shift, when average pairwise correlation jumped from 0.25 to 0.5 and 252-day tracking correlation fell from 0.945 to 0.413; measured the stability vs adaptability trade-off across estimation windows (turnover 99% at 21 days vs 73% at 252 days) and extended to the Russell 1000/2000 (tracking-error vol 9.8% S&P 500, 11.0% R1000, 22.9% R2000)

Feb 2026 – Present

• Python event-driven backtest framework: walk-forward CV, block bootstrap, BH-FDR correction; 90-config sweep across detectors, hold periods, and HMM regimes on Russell 1000 tick + order-flow data
• Extended to prediction markets via Kalshi + Polymarket capture — 18,000+ subscriptions at ~1,100 msg/sec, 14M+ events across sports and event contracts for signal generation

A graduate research project (MF821) that fits a VAR(1) cointegration model on intraday mid-prices for five sector-diversified pairs, extracts the cointegration factor via spectral decomposition, and solves free-boundary PDEs by finite differences to derive time-dependent optimal entry/exit bands. Backtested out-of-sample against ad-hoc σ-bands and Bollinger heuristics across 200+ trading days using real NBBO spread costs.

Sept – Dec 2025

An LLM-driven sentiment-to-signal framework for equities that scores news and filings with FinBERT, layers in cross-source disagreement features, and converts them to z-score thresholds for entry/exit. Designed as a reproducible out-of-sample evaluation pipeline rather than a curve-fit backtest.

Sept 2025 – June 2026

• Modeled salary vs fWAR across 4,246 player seasons in Python; regression residuals flag over/underpaid contracts, positional scarcity, age curves, and development efficiency
• Built end-to-end on Next.js, FastAPI, and Cloudflare R2 with a 14-dimension team contention model and roster simulator

Apr 2026

• Four-stage pricing pipeline (MF728, three-person team) for a 1Y×5Y ATM payer Bermudan on $10MM notional, benchmarked against Bloomberg’s HW1F NPV of $235,823: SOFR curve → SABR calibration → LMM simulation → Longstaff–Schwartz Monte Carlo
• Owned the SOFR curve: bootstrapped discount factors from convexity-adjusted SOFR futures and 1Y–50Y OIS swaps with cubic-spline zero rates, matching Bloomberg’s zero curve exactly; SABR (β = 0.5, Hagan 2002) fit Bloomberg VCUB normal vols across 98 expiry-tenor pairs at 0.38 bp mean smile error
• 50K-path antithetic LMM reprices the OIS curve within 0.03%; LSMC prices the Bermudan at $213,580 (100K paths, s.e. $844) — the European leg lands within 0.55% of Bloomberg, and the 9.4% Bermudan gap reflects the LMM vs HW1F model-class difference; sensitivities to correlation, vol, curve shape, SABR ρ and path count all move with the correct sign

• Undergraduate thesis (four-person team) wrapping a pre-trained YOLOv8m object detector in an explainability layer, applying SHAP and LIME to show which image regions drive each detection
• Built a LIME adapter for object detection (custom predict_proba over detections) and ran explanations on street-scene, licence-plate and live-sports footage toward a litter-detection use case; found LIME’s local explanations too coarse for detection models and moved to SHAP attributions
• Extended the pipeline to medical imaging in collaboration with Princess Margaret Hospital (results confidential)