AI strategist. Inventor.
Snow skier. Vibe Coder.
Not necessarily in that order.

Overview

Adidas wanted a finish-line activation at the 2024 City2Surf: racegoers answer a trivia question, get it right, and a box of shoes lights up. With five days on the clock and no prior app-building experience, I architected, designed, built, and tested the entire solution myself — a touchscreen quiz on a Commbox display, a Python app integrated with API-controlled lights, all running off a dedicated WiFi box to sidestep the 4G chaos that comes with 90,000 people crossing a finish line.

My Role

End-to-end ownership. I figured out what hardware was needed, connected the quiz logic to the lights API, and made it robust enough for real people in a loud, high-stakes environment. Every architectural decision and every line of code — vibe-coded with ChatGPT and pasted into VS Code, well before tools like Cursor or Claude Code existed — was mine.

Outcomes

• 500+ racegoers interacted with the activation at the finish line
• Vibe-coded and production-ready in 5 days, with no prior app-building experience

Lessons Learnt

At mass-participation events, 4G/5G is essentially unusable — 90,000 registered participants at a finish line will saturate any mobile network. In hindsight, a Raspberry Pi running locally would have been far more resilient than relying on WiFi. It's the kind of thing you only learn by shipping something real.

Overview

Ask Johnny started as a personal learning project — I wanted to understand how retrieval-augmented generation (RAG) actually works by building one end-to-end. I chose Ausplow as the domain because the subject matter is highly specific: deep ripping, seeding systems, soil constraints, and machinery setup aren't things a general-purpose AI answers well. That made it a good test of whether a grounded retrieval system could outperform a generic one on niche, technical content.

My Role

I designed and built the entire proof of concept myself using Python, Streamlit, OpenAI APIs, ChromaDB, and CrewAI — crawling and ingesting the Ausplow website and PDFs, converting them into embeddings for semantic search, architecting the retrieval pipeline, and building a chatbot interface to query it.

Outcomes

• Successfully built a working RAG pipeline end-to-end — from raw web and PDF content through to grounded AI responses
• Came away with a hands-on understanding of vector databases, embedding strategies, and retrieval architecture

Lessons Learnt

The hardest part wasn't generating answers — it was data quality. Cleaning duplicated website content, structuring documents properly, and ensuring the retrieval layer surfaced the right context before the model responded were what made or broke the system. The limiting factor in RAG isn't the AI — it's the pipeline that feeds it.

Overview

Quoting is one of the most time-consuming and inconsistent parts of running a trade business. Most builders are doing it manually in Word or on paper, with no standard structure and no easy way to scope labour, materials, and exclusions accurately. The Quote Builder is a web app that guides tradies through a multi-step AI chat experience — entering job details, scoping work per trade, calculating costs, and exporting a professional quote as a Word document. A pricing engine handles the numbers (rule-based, not AI — intentional for accuracy), while Claude handles the language: scope descriptions, cover letters, and exclusions. A one-off Stripe payment unlocks the download.

My Role

I've been driving the end-to-end product build — from the backend API routes in FastAPI and Supabase, through to the document generation pipeline using python-docx, Stripe payment integration, and the AI prompt architecture across the six Claude calls that power the flow. I also reviewed the NSW Government home building contract template for compliance requirements, produced the full E2E flow document for designer handoff, and created the system architecture diagram for the GitHub README.

Features

• AI-guided chat flow that extracts job details, client info, and site scope from natural language input
• Rule-based pricing engine for accurate labour and materials calculations
• Live Bunnings price lookups via a carpentry materials calculator
• AI-generated scope descriptions, cover letter, and exclusions tailored to the job
• Professional Word document export built with python-docx
• One-off Stripe payment per quote — no subscription required

Outcomes

• MVP built and being used by a small group of builders today
• Reduces quote turnaround from hours to minutes for a trade business

Lessons Learnt

Keeping the pricing engine rule-based rather than AI-generated was the right call — builders need to trust the numbers, and AI-generated costs introduce variability that erodes that trust quickly. The more interesting design challenge was the chat interface itself: making it feel like a natural conversation while still capturing structured data accurately enough to power a compliant, professional document at the other end.

Overview

With AI moving fast, I wanted a way to keep my team informed without anyone manually trawling through the news each morning. So, I built a Python script that pulled from NewsAPI, used GPT-4 to summarise and categorise the day's top AI stories into four sections (trends, markets, policy, and innovation) and pushed it to a Slack channel every day at 9:40am. Each summary was saved locally so the next day's digest could be diffed against it to avoid repetition.

My Role

I designed and built the full pipeline myself, which included API integrations, the GPT-4 summarisation prompt, the Slack webhook, local file storage for deduplication, and the scheduling logic. It ran autonomously once set up.

Outcomes

• 50 team members joined the Slack channel organically within 2 months
• Delivered a daily digest autonomously with no manual effort after initial setup

Lessons Learnt

NewsAPI's free tier has rate limits that occasionally clipped coverage; a multi-source feed would make it more robust.