I build AI-powered products end to end.

An iOS app for friend groups who run a weekly pub night. It handles whose turn it is to pick, voting, ratings, leaderboards and reminders — and recommends new pubs based on what the group has loved. Behind it, an automated pipeline reads 507 Sydney venue websites, uses AI to extract each pub's weekly specials, and routes anything uncertain to a human for approval. The result: a live deals catalog with no content team.

582

APPROVED DEALS

124

VENUES COVERED

204

PIPELINE RUNS · 0 FAILED

0.93

AVG AI CONFIDENCE (0–1)

Rules first, AI second, a human sign-off last.

1. STEP 01

   Start cheap and predictable

   Scheduled jobs visit each pub's own website and menu PDFs — never third-party listings, which testing showed can sit years out of date. Simple pattern-matching trims each page down to the handful of lines that look like deals, so most of the work costs nothing to run.

   507 VENUES CHECKED AUTOMATICALLY
2. STEP 02

   AI with guardrails

   Google's Gemini model turns those snippets into structured deals — what, which day, what price — with a confidence score on every one. Duplicates are removed, one-off events are filtered out, and results are cached for 30 days so the AI is never called twice for the same venue.

   0.93 AVG CONFIDENCE
3. STEP 03

   A human has the final say

   Anything the AI is less than sure about lands in a review queue inside the app, and only approved deals ever reach users. Every published deal links back to its source and shows when it was last verified — so nothing stale or invented makes it to the app.

   204 RUNS · 0 FAILED

UNDER THE HOOD: DENO EDGE FUNCTIONS ON SUPABASE · SSRF-SAFE FETCHING · PROMPT-INJECTION DEFENCES · PER-USER RATE LIMITS · POSTGRES ROW-LEVEL SECURITY · EVERY RUN LOGGED · 109 MIGRATIONS · 320+ AUTOMATED TESTS

• −25%

  Key iOS contributor on the bank-grade login migration (ForgeRock). Login-related calls to the contact centre dropped 25% after launch.

• LEAD

  Led the mobile front-end migration for Suncorp's highest-priority program — Duck Creek insurance data — defining shared patterns across iOS and Android.

• −50%

  Shipped the self-service features that contributed to call-centre contact rates falling more than 50% over five years.