👋 Hi, I'm Yu-Wei

Selling hotel rooms comes with work the traveler never sees: screening for fraud, confirming with the hotel, answering the guest, refunding, and reconciling with suppliers. At Pilot that work was spread across seven tools, including payment and support dashboards, a shared inbox, Slack alerts, and spreadsheets.

I started Control Tower as a prototype on mock data shaped like our production database, and had it deployed within four days. Now one page shows everything about a booking, with an audit log that names who changed what. ⌘K finds a booking by any of nine identifiers.

Recurring chores became worklists. Hotel confirmation calls run from a keyboard-driven queue that loads the whole week, so nobody changes pages mid-call. Supplier reconciliation went from a hand-filled spreadsheet to a page where ops paste a statement and see what matches. Bans got a console with reasons, expiry, and an audit trail. Engineers on the team added cancellations, refunds, and booking recovery.

Watchtower is Pilot's fraud engine. It scores every booking during checkout, then approves it, asks for stronger authentication, holds it for a person, or denies it.

Fraud rules can fail two ways. Too strict, and good travelers get turned away. Too loose, and chargebacks get through. I audited 90 days of Watchtower's production decisions, tested each signal against what it claimed to measure, and fixed or retired the ones that didn't hold up.

Then I rewrote the decision policy around one rule: deny only on clear evidence, and send anything uncertain to a person, because a hold can be undone and a denial can't. Replayed against the same 90 days, the new policy turned away no good bookings and sent every known fraud case to review.

I also added a second, LLM-based check that runs in shadow mode. I chose its model with an eval that ranks harmful mistakes above accuracy, speed, and cost. Gemini 3.5 Flash-Lite got all 90 cases right. Claude Haiku 4.5 got 87, and all three misses were the harmful kind.