Builds I’m proud of: a fraud-detection pipeline, a multiplayer park, a plain-English SQL tool, a local-first desktop app, forecasting models, and a few smaller tools. Each one runs end to end, from source data to something a person can use.
Featured
The three I’d show first
Latest · LLM + SQL in the browser
2026
Quack Query Live
Ask a data file questions in plain English: an LLM writes one read-only SQL query and DuckDB runs it in your browser.
Asking a data file a question usually means writing SQL, or uploading the file to someone else’s server.
Fix
An LLM writes one read-only SELECT from the table profiles, and DuckDB runs it in the browser, so the file never leaves the device.
Result
Live. Every answer shows its SQL, which you can edit; read-only checks run on both server and browser, and columns that look personal are kept from the model.
Data engineering · Fraud analytics · Chapter 2
2026
Fraud Detection Pipeline Running
Chapter two: a new real-time model scores every simulated webstore checkout as it lands in Supabase, and caught 110 of 111 fraudulent orders in its first 24 hours live, Nov 14 to 15. Chapter one, the baseline rules it now runs beside, flagged about half the fraud and no legitimate orders.
Five fraud patterns hide in about 3% of webstore orders, and the true label only arrives 5 to 60 days later, the way a chargeback does.
Fix
Chapter two, since Nov 14: a new real-time model scores every checkout as it arrives and logs its fraud probability and scoring time to its own table, beside the rules, so the two are measured on the same orders. Chapter one is those rules: weighted rules score each checkout in time order and give reason codes, and a score that takes over 150 ms fails open, so a real guest is never left waiting.
Result
In its first 24 hours live (orders placed from 5:27 PM ET on Nov 14 to the same time on Nov 15), the new model caught 110 of 111 fraudulent orders where the rules caught 52, at a cost of 25 good orders flagged out of 1,964. Before it, the baseline rules flagged about half the fraud and no legitimate orders, measured against the late labels.
New model, first 24 hours110 of 111 caught
Rules, fraud flaggedabout half
Rules, good orders declined0
Product analytics · Multiplayer web app
2026
Owl Park Live
A live multiplayer owl park with Shopify tickets; every visit and sale lands in Supabase for analysis.
Sales and visits are only worth analyzing if every event lands exactly once, even when a send fails.
Fix
Signed Shopify webhooks issue the tickets. Each event goes to an outbox in the same step as the action it records, then ships in batches, and event ids make a resend land once.
Result
Visits and sales arrive in a type 2 player dimension and fact tables, with views for minutes per room, free-to-paid conversion and lifetime value.
SQL source
SELECT title, badge, eyebrow, year, summary, tags
FROM projects
WHERE featured
ORDER BY sort_order;
A local-first desktop app that answers questions about a small group of files, with citations back to the passages it read. A Rust engine in a Tauri 2 shell handles extraction, on-device OCR and embeddings, and hybrid retrieval. A question over a spreadsheet becomes one read-only SQL query, run by DataFusion and shown verbatim. Use the bundled private model or bring your own key. Installers for Windows, macOS, and Linux.
Getting answers from a small set of private files without sending them to a server.
Fix
A Rust engine handles extraction, on-device OCR, embeddings and hybrid retrieval. A question about a spreadsheet becomes one read-only SQL query, run by DataFusion and shown verbatim.
Result
Version 0.15 ships installers for Windows, macOS and Linux, and every answer cites the passage it came from.
Forecasting · Machine learning
2026
Attendance Forecasting
Daily theme-park attendance forecast from weather, calendar, and flight-traffic features. A blend of gradient-boosted trees, ridge, and random forest selected by rolling-origin cross-validation beat a tuned XGBoost by about 6% MAE out of fold and cut mean bias from +90 to −7 guests a day. The last 30 operating days are held out as an unseen live forecast with an 80% planning band from pooled out-of-fold ratios.
A tuned XGBoost model was the daily attendance forecast to beat.
Fix
A blend of gradient-boosted trees, ridge and random forest, chosen by rolling-origin cross-validation.
Result
About 6% lower MAE out of fold than the tuned XGBoost. The last 30 operating days are held out as an unseen live forecast with an 80% planning band.
Mean bias, guests a day+90 → −7
Forecasting · Machine learning
2026
USL Attendance Forecasting
An XGBoost demand model on nine seasons of USL Championship matches, on a staged DuckDB warehouse (raw → staging → marts). Two models share every upstream table and differ only by league-position features, so the error gap measures what promotion and relegation will do to attendance from 2028.
USL promotion and relegation start in 2028, and there is no data yet on what they do to attendance.
Fix
Two XGBoost models share every upstream table on a staged DuckDB warehouse and differ only by league-position features.
Result
The error gap between the two models is the measuring stick, set as a baseline before promotion and relegation begin.
Data quality · Pipeline
2026
Roll Call
A daily manatee-count pipeline for Blue Spring State Park, built around its data-quality layer: freshness, volume, distribution, and cross-source checks that quarantine suspect observations, open incidents, and email what newly broke. DuckDB on a systemd timer; an Evidence dashboard on Vercel.
A daily manatee count is only useful if one bad observation can’t quietly skew it.
Fix
Freshness, volume, distribution and cross-source checks quarantine suspect observations and open an incident.
Result
It runs daily on a systemd timer, emails what newly broke, and publishes the count to an Evidence dashboard.
Learning project · Data assimilation
2026
Attendance Nowcast
A Kalman-filter nowcast of daily theme-park attendance: predict from yesterday’s state, then correct with today’s partial signals, weighting each by its uncertainty. Written up as a five-notebook series from the scalar filter upward.
A forecast made yesterday ignores what today’s partial signals already say.
Fix
A Kalman filter predicts from yesterday’s state, then corrects with today’s signals, weighting each one by its uncertainty.
Result
A five-notebook write-up that builds the filter up from the scalar case.
Analytics · Power BI
2025
Member Segmentation (RFM)
Recency, frequency, and monetary segmentation of membership data, modeled in SQL and reported in Power BI, to turn one big member list into groups a marketing team can act on differently.
Recency, frequency and monetary scores modeled in SQL.
Result
Member groups, reported in Power BI, that a marketing team can treat differently.
Machine learning · Audio
2026
eFrog Retired
An audio-classification model that identified 19 Floridian frog species from a recording. The web app is retired; the model and code are still on GitHub.
The first Owl Park: a simulated ticket-sales ecosystem where n8n agents role-played customers and managed stock, Supabase stored the orders, Microsoft Fabric modeled them medallion-style, and Power BI turned them into decisions. Retired, but the architecture diagram is still worth a look.
More, including learning repos for dbt, Dagster, Docker, and assembly, on GitHub →
How the data flows
Tap a stage to see what happens there and what runs it
Fraud Detection Pipeline
One run of the generator, from seed tables to raw tables in Supabase, scored by the rules and, since Nov 14, by a new real-time model too.
1 · Source Seed tables
Products, a pool of real guest profiles and a separate pool of fraudster identities, each tagged with the fraud pattern it plays, sit in Databricks seed tables. A Python script makes the identity pools; each run reads them once.
Databricks
Python
2 · Ingest Simulated checkouts
A Databricks notebook job simulates a window of Owl Park webstore traffic, by default an hour: about 120 normal orders plus fraud episodes worth about 3% of orders. The five patterns are card testing, stolen-card resale, gift-card cash-out, account takeover and fraud rings.
Databricks Jobs
Python
3 · Transform Order features
Orders are scored in time order so velocity counts are honest. Each one gets features such as cards seen on its device in the last hour, a headless browser, the gift-card share of the cart, an IP and billing country mismatch, and the account's age.
Python
4 · Score Rules model and checkout
Weighted rules turn the features into a 0 to 1 score with reason codes: approve, send to an identity check, or decline. The checkout is simulated too, with issuer declines and latency. If scoring takes over 150 ms the order fails open and is approved, so real guests are never stuck waiting.
Python
Rules model v0
5 · Load Four raw tables
Transactions, fraud decisions, checkout responses and fraud truth land in Supabase in one transaction, so a run arrives whole or not at all. Every column except the id and timestamp is nullable, so a partial row is kept for inspection instead of lost, and row-level security keeps the tables private.
PostgreSQL (Supabase)
psycopg2
6 · Score again New real-time model
Since Nov 14 a new model scores each checkout as it lands, beside the rules. For every order it writes a fraud probability, a flag and how long the scoring took to its own Supabase table, realtime_fraud_audit, keyed by the same transaction id, so both scorers can be judged on the same orders.
PostgreSQL (Supabase)
7 · Label Late fraud labels
Each order's truth row carries its real fraud flag, but the label only becomes available 5 to 60 days later, and only for fraud that got through, the way chargebacks do. With decisions, checkout responses and labels side by side, a query can measure the trade-off for each scorer: how much fraud was flagged, how many good guests were turned away, and how many orders failed open. The live page compares the two.
From a ticket sale to the analytics views, as the park runs today.
1 · Source Shopify store
Day Tickets and Annual Passes sell through a Shopify store. The buyer checks out in the store's own window, and Shopify keeps the buyer's details.
Shopify
2 · Ingest Signed webhooks
Shopify posts each paid order to a Cloudflare Worker, which checks the webhook's signature before reading anything, so a forged order never reaches the tickets, and a repeated webhook changes nothing. It then issues one ticket per unit into a ticket book kept in a Durable Object.
PlannedScore every paid order with the fraud pipeline before its tickets release, waiting up to 2 seconds and failing open.
Cloudflare Workers
Durable Objects (SQLite)
3 · Capture The park itself
Visitors walk the rooms as owls. An authoritative room server keeps every owl in sync over hibernating WebSockets, and the turnstiles let an owl in once its ticket scans. Visits, room stays, collectibles and ticket events all become rows.
Phaser 3
WebSockets
4 · Load Outbox to Supabase
Each event is written to an outbox in the same step as the action it records, then sent in batches. A failed send retries later, and event ids make a row sent twice land once. Visitors leave as salted hashes, never names or emails.
Durable Objects
Supabase Edge Functions
5 · Model Analytics schema
Rows land in a game schema shaped for analysis: a type 2 player dimension, where each change to an owl's kind or color opens a new version, beside session, room-visit and purchase fact tables.
PostgreSQL (Supabase)
SQL
6 · Serve Views for product questions
SQL views answer the questions a park owner asks: minutes and bounce rate per room, free versus paid visitors and when a free visitor converts, each order's status, and lifetime value per owl.