Human performance,
quantified

Sprint Performance Projection Dashboard

Problem: Coaches and athletes have no principled way to project next-year sprint times — most rely on intuition or flat conversion tables that ignore individual development trajectories, event profiles, or the structural difference between same-season equivalency and year-over-year growth.

Built a single-file Dash app backed by a SQLite database of Section V meet results spanning five sprint events (55m, 100m, 200m, 300m, 400m). Implemented two model families: same-season equivalency models (OLS + Random Forest, 630 models across events, sexes, and grade levels) to contextualize a current performance relative to peers; and next-year projection models (Mixed Linear Models with random athlete intercepts, ~300 models) to forecast next season's best time from this season's trajectory. Season features include best time, first-meet time, improvement rate, and linear and quadratic trend. Validated via temporal holdout — trained on transitions through 2023, tested on 2024→2025. Integrated the Claude API to generate coach-facing narrative interpretations with SHA-256 response caching and a constrained system prompt that classifies model type, flags sparse-input projections, and enforces domain-appropriate language.

Role: Sole developer and analyst

Proprietary

Force Plate Assessment Reports

Problem: Coaches and clinicians collecting force plate data had no efficient way to turn raw signals into interpretable, branded reports they could use with athletes and patients.

Built an end-to-end system that pulls raw force-time data via the Hawkin Dynamics API, applies signal processing (Butterworth filtering, COP centering, Savitzky-Golay differentiation), computes derived metrics (sway path length, 95% confidence ellipse, asymmetry indices, EUR, DSI), and generates multi-page PDF reports with trend analysis across sessions. A Flask interface lets non-technical users select athletes, configure tests, and generate reports without touching code.

Role: Sole developer and analyst

Gait Deterioration & Fall Risk Classification

Problem: Standard gait assessments capture averages across an entire walk — but fall risk may be better predicted by how gait breaks down over time, especially under fatigue.

Designed a quarters-based segmentation approach for 6-minute walk test data from wearable IMU sensors (APDM Opal). Instead of averaging across the full walk, I extracted gait features per quarter to capture fatigue-related deterioration patterns — speed decline, variability changes, asymmetry shifts. Built the full analytical pipeline: feature engineering, assumption testing, effect size ranking, FDR correction, and logistic regression with ROC/AUC evaluation.

Role: Sole analyst and pipeline developer

Machine Learning for Fall & Dementia Risk

Problem: A clinical gait and mobility assessment produces hundreds of candidate measures. Which of them actually carry signal about fall history — and can they classify risk well enough to matter clinically?

Principal-investigator project at the University of New Hampshire. Screened high-dimensional gait and mobility features with false-discovery-rate adjustment to separate signal from noise, then compared regularized linear models against tree-based ensembles for fall-history classification, evaluated with ROC/AUC.

Role: Principal investigator · sole analyst

Hand-Timing Bias in Section V Sprint Results

Problem: High school sprint times are routinely recorded by hand — but hand timing introduces a systematic positive bias relative to fully automatic timing (FAT). How large is that bias in practice, does it vary by performance level, and does it differ between the 100m and 200m?

Scraped and cleaned 17 years of Section V outdoor 100m and 200m results. Applied a within-athlete paired difference design — comparing each athlete's best hand-timed season performance against their best FAT performance in the same season, within a fixed date window that excludes championship meets. Binned results by FAT speed tier to quantify how bias magnitude varies across the performance distribution. Conducted the analysis separately by sex and event, then cross-compared 100m vs. 200m findings.

Role: Sole developer and analyst

MLB Baserunning Workload Prototype

Problem: Baserunning is repeated near-maximal sprint work, but it is rarely quantified as workload. Can player-tracking data support exposure and recovery metrics for baserunners?

Prototype built on MLB player-tracking data (2026): extracted peak 1-second speed per baserunning opportunity, then constructed exposure counts and recovery windows across games to characterize sprint load at the player level — the same load-monitoring logic used in track and field, applied to a tracking-data stream.

Role: Sole developer and analyst

Hawkin Dynamics API → Structured Data Pipeline

Problem: Organizations using Hawkin Dynamics force plates had athlete data locked behind an API with no easy way to extract, organize, or analyze it at scale.

Built Python scripts to extract full organizational data from the Hawkin Dynamics API — athletes, teams, groups, and all historical assessment data. Handles pagination, rate limiting, and data type mapping, then outputs clean, analysis-ready Excel workbooks.

Role: Sole developer

VR-Based Balance & Cognition Intervention

Problem: Can immersive VR training improve balance and cognitive function in older adults? Answering this requires a tightly controlled longitudinal study with a population that's hard to recruit and retain.

Designed and executed a multi-session intervention study (2x/week over 8 weeks) with older adult participants. Managed the full research operation: IRB protocols, informed consent workflows, equipment calibration, session-level data collection, and participant scheduling across repeated lab visits. Integrated cognitive, balance, and self-report datasets across timepoints to assess intervention effects and practice-related changes.

Role: Study lead — design, operations, and analysis