I turn complex ideas into working software: sports prediction systems, lineup optimizers, and the workflows that keep them running. I lead the design, architecture, testing and operation, and use AI coding agents to build and refine the tools.
11 on view, 8 live: a public picks board, 4 prediction back ends, an NFL lineup optimizer in season, a bet tracker, and this site’s own article pipeline. 2 retired models keep their numbers; one optimizer is between seasons.
Each project is built with AI coding agents working under a written procedure: separate agents plan, build, test and review, a reviewer cannot change code, and I approve what ships.
A public site showing model picks for NHL, MLB, NFL and six soccer leagues, read straight from four prediction databases. Every pick is graded in public: accuracy, Brier score and log loss per sport, in probability buckets. Edge, expected value and ROI stay hidden until audited. Live since 2026.
A web app for GameBlazers NFL card contests. Upload a card roster, choose projections and contests, and it solves one integer program per lineup, in your priority order, from a shared pool of cards under salary caps, position slots, locks, stacks and exposure limits. Then it manages the week: entered lineups, kickoff locks, late swaps. Live since August 2026.
Four Python back ends, one per sport, each ingesting its own data, training walk-forward, and writing calibrated probabilities to a database that sharps.gg reads directly: the database views are the contract. Each runs on its own scheduler with failure alerts.
Power ratings fit walk-forward over 27 seasons; recipes frozen and pre-registered before results are read.
Stacked gradient-boosted game models with weekly gated retrains; a database-owned switch changes the displayed model without a deploy.
A stacked ensemble over 17 seasons; odds kept as a pre-game time series after an audit found most stored closing lines were in-play.
Six leagues, a gradient-boosted match-winner model nightly; 15,728 picks graded in public at a hit rate matching the stated probability, as of 2026-10-25.
A personal bet-tracking app. Screenshot a sportsbook slip, a vision model drafts the bet, you confirm it, and an hourly job grades each leg against real scores and records the closing line. Every dollar shown is summed from an append-only ledger, so the ROI figures cannot drift from the bookkeeping.
Essays for smart readers who do not write code, about what AI actually does: to jobs, to insurance claims, to the betting line. Newest first.
Entries come from Ghost at request time, newest first. All entries.
RSS feed