Projects

• Good360AI and Data engineering for a $3.9B nonprofit

  

  AI agents in production

  Built a governed semantic layer, 3 MCP servers and 5 AI agents, unlocking self-serve AI analytics, automating project management, outperforming vendor disaster monitoring, and taking Power BI documentation from weeks to minutes.

  

  Warehouse ownership

  Own and operate a Synapse warehouse and its ingestion pipelines across 10 source systems and hundreds of millions of rows, and authored the curated SQL layer the business reports from.

  

  Major data integrations

  Built idempotent Salesforce delta loads into the data warehouse, integrating 60% of donations with no automated visibility, adding $1.42B in recipient attribution and raising 990 Schedule I coverage from 53% to 94%.

  

  Web portal for impact reporting

  Launched a live donor portal that serves major corporate partners automated impact reports, pushing the boundaries of nonprofit data.

  

  Self-serve analytics

  Designed a hub-and-spoke semantic model and a complete analytics and impact reporting platform, scaled from zero to 1,000+ monthly views across 20 departments.

  

  $1.5M in grant funding

  Lead data strategy across all grant reporting, and secured $1.5M in grants funding my team's tech development, from the proof of concept through renewal.

  Azure Synapse, T-SQL, Power BI, DAX, Python, FastAPI, MCP, Claude agents

• USD BaseballLeading Analytics work for WCC Champs

  

  Built for and with coaches

  Shipped apps for game prep, pitch design, player evaluation, and scouting.

  

  Custom Player Evaluation Models

  Pitch-level Stuff+, xWhiff, xSLG and run value models.

  

  Automated Transfer Portal Dashboard

  AI research automation paired with position-relative 20-80 tool grades to find the strongest candidates we can realistically land.

  

  Predicting arm angle from release point

  3.6° RMSE on hold-out set, used to estimate arm angle where data is unavailable to let hitters visualize a pitchers arm slot.

  Python, TrackMan, XGBoost, LightGBM, Streamlit, dbt, DuckDB, Supabase

• PRTLA solo startup building AI art frames and their mobile app

  

  A native app on Expo

  An Expo and React Native app drives every frame through a Cloud Run API, with Firestore pushing each change to the wall in real time.

  

  AI art, made for each frame

  Gemini 3 Pro Image renders 46 art styles at 2K, giving every frame on the wall its own variation.

  

  Live data rendered as art

  The market, weather, sports and commute modes turn live data into wall art, including stock charts rendered in Python.

  

  Seventeen live modes

  Gemini-generated and server-rendered SVG modes turn art, weather, markets, calendars and clocks into decor.

  

  Every frame generated in parallel

  The art modes fire one Gemini call per selected frame in parallel, so a whole home refreshes at once and no two walls match.

  Expo, React Native, Node, Cloud Run, Firestore, Gemini

• DocPBISaaS that reverse-engineers Power BI files into docs

  

  Pre-launch

  Software that documents Power BI reports for you. Built, priced, and launching soon.

  Gives line of sight to any audience

  Unzips the report, decompresses its XPress9 data model, and reads the SQLite metadata to write business and developer docs.

  

  Automated version notes

  Compares two uploads of a report and lists the table, measure, and visual changes in plain English.

  

  Many reports, one model

  Matches reports to the shared model they run on using the dataset ID stored in each file.

  Python, FastAPI, Celery, Postgres, Redis, React

• Under ParSalary cap based fantasy golf web app

  Fantasy golf made easy

  Start a league, invite friends by code, draft golfers under a $300 cap, and follow the season standings.

  Marketing built with code

  The promo is React code, rendered to video by Remotion and built with Claude Code.

  

  Automated Data feeding hole by hole scoring

  Automated data pipelines for betting-market odds to set salaries, and hole by hole scoring to keep users in the action.

  Next.js, Prisma, Cloudflare Workers, Remotion

• Fantasy FootballLeague hub, trade simulator and NFL stats web app

  

  One hub for every league

  Merges Sleeper and ESPN APIs, then ages dynasty rosters 5 years out.

  

  Trades both sides take

  Screens 4,600+ packages, then Monte Carlo title odds for both teams.

  

  Play-by-play research lab

  Targeted-route trees and man vs zone splits from nflverse charting.

  

  Kicker rankings that beat Vegas totals

  Ridge on 16 features out-ranks Vegas implied totals (p = 0.001).

  Vanilla JS, Python, scikit-learn, nflverse

• Sports ModelingData science for sports

  

  Bracket forecasting

  Ridge regression turns 16 Bart Torvik team stats into a strength score, and 10,000 simulated brackets turn it into odds by round. A pre-registered test on 1,007 games found no edge over seed odds.

  

  12 Days of Madness

  A 12-day X series, 136 posts and 93 charts, on what predicts a March run, later rerun with bootstrap intervals and a multiple-testing correction: 11 Bart Torvik metrics against how far 1,152 tournament teams went. Most fade within a seed line, schedule strength from 0.47 to 0.07; offensive efficiency is the one that survives the correction.

  

  Game-by-game baseball models

  Rebuilt a 604-line MLB run-expectancy script into a tested R pipeline, then ran a pre-registered study of which look-back window to trust for hitters, starters and bullpens; the model it produced called 58.0% of 12,148 games from 2021 to 2025 at first pitch, beating a run-differential model in all five seasons. Then built a plate-appearance matchup model (log5 batter-pitcher rates, park, times through the order, bullpen) and a Monte Carlo game simulator; with the starter guessed from the rotation, its 955 picks at the opening betting line from 2023 to 2025 gained 0.7 points of closing-line value each, about twice what betting every home team drew. Separately scored 0.726 AUC on the private test set of Nick Wan's batted-ball Kaggle competition.

  Python, R, scikit-learn, CatBoost, testthat

• BoondockOpen source app that finds free camping on public land

  

  Google Earth, for campers

  Scout any campsite or trail in 3D satellite before driving out.

  

  Can I camp here?

  Tap any spot for a green, yellow or red answer citing the land manager, nearest open road and stay limit, plus corridors marking where camping is likely legal.

  

  Offroad drive time visibility

  Routes where Google and Apple Maps stop, timing paved and dirt legs.

  TypeScript, MapLibre, Turf, ArcGIS REST

• Built For MyselfPersonal apps to maximize agency

  

  Personal Finance App

  A custom personal financial and budgeting app pulling data through Plaid, a hand-built Webull client and a shared budget sheet every night.

  

  A stock screener built from raw SEC filings

  A DuckDB warehouse of 9.18M XBRL fundamentals rows across 6,732 companies feeds the factor model, backtests and options math.

  

  Scraping market sentiment reliably

  An LLM filters and grades X, Reddit and StockTwits chatter (0.90 AUC, beating a model trained on 27.8k labels), feeding a screen that ranks stocks, options and prediction markets by upside.

  

  Data-driven personal health

  Ranks six Apple Health metrics against published population norms, and scores each day 0 to 100 from seven signals against my own baseline, naming the biggest driver.

  

  Making apartment searching objective

  Each roommate ranks their feature priorities, then every toured apartment gets graded and plotted against the score its rent has to clear.

  Python, DuckDB, HTML, Cloudflare Pages