Passion Projects
Side projects built purely because they were fun to build.
Phillies Prospect Pulse
A self-updating farm-system dashboard with a custom prospect ranking model and a Claude-powered chat assistant that reasons over the live database.
An automated pipeline pulls and reconciles news, stats, promotions, and injuries on a schedule, feeding a ranking model I designed from scratch โ weighing scouting, performance, age/level, media sentiment, and risk โ that shows exactly what moved each player's rank. "Prospect Genie" is a live, Claude-powered chat assistant that reasons over the database to answer natural-language questions about the system โ why a player hasn't been promoted, who's trending, head-to-head comparisons. Uses Supabase, and an automated quality pipeline (data audits, ranking backtests against historical outcomes, and typechecking) that runs before every deploy.
Opportunities to learn
- Having weighed Claude against DeepSeek and Llama for this Claude assistant โ tested live against the real API, not just compared on paper โ a future project could build a provider-switchable AI layer to compare cost, latency, and quality head-to-head.
- Now that a full backtesting pipeline exists, a future project could swap the weighted ranking formula for a trained model and compare the two
MycoFlow
A mushroom cultivation tracker built on a real phase state machine โ every batch followed from grain jar to dried harvest, with an audit trail that can't drift.
Every batch moves through grain colonization, break & shake, bulk colonizing, fruiting, drying, and done, with location assignment and dried-weight yields logged at each step. The lab map (shelves and chambers) derives occupancy from each batch's location instead of storing the relationship twice, and every mutating action writes one entry to both the batch's own history and a global activity log in the same call โ so the audit trail can't drift out of sync. Next.js + TypeScript on the frontend, Postgres (Neon) on the backend.
Opportunities to learn
- Having modeled a real phase state machine here, that same pattern is reusable for any workflow-tracking app, not just cultivation
- Now that the audit-log pattern (one write, two tables, always in sync) is proven out, future projects can reuse it anywhere mutations need a paper trail