AI News and Shipping a Paid AI Car Recommendation Platform — 02Ship Sydney Meetup — October 20, 2026
AI News and Shipping a Paid AI Car Recommendation Platform
TL;DR
- Use the LLM to understand the user, not to make the final pick. Keep the recommendation deterministic.
- AI can fill data gaps, but a human admin approves new data before users see it.
- Build a golden prompt set and rerun it every time you change the model or the prompt.
- AI is overly agreeable and has no common sense without context. Tell it about concurrency and throughput upfront.
- With AI coding, a job estimated at 28 weeks shipped in about 12 weeks.
On Tuesday, October 20, builders came to Haymarket HQ for the October 02Ship Sydney meetup.
The night had a monthly AI news update, a real client project shipped with Claude, and a long round of community shout-outs.

Thanks to Haymarket HQ for hosting us again.
Talk 1: AI Monthly Update — Bob Jiang (02Ship)
I opened with a quick round of AI news from the past month.
OpenAI
- o3 — capable at personal agent work.
- o4-mini — flagship-level capability at a lower price.
- UltraFast — a faster processing tier inside ChatGPT. It costs more. It is not a model upgrade.
- Responses API — lets your app route decisions, similar to function calling.
- CodeSandbox Cloud — background coding in the cloud. You come back later and review the work.
- Intelligent UI — interactive, non-text answers. Rolled out on October 7 for paid users and October 8 for free users.
Anthropic
- The new Claude 5.5 family: Haiku for fast and scalable work, Sonnet for product development, Opus for complex tasks.
- Claude Code Modes — custom panels, rules, and reusable workflows you can share with your team.
- Claude-managed agents — scheduled, recurring tasks that run in the cloud.
Personal AI assistants
I grouped them into three types:
- Work-focused assistants.
- Personal assistants for email, subscriptions, and daily coordination: Muse, Instinct, Epoch.
- Home automation: Ball, Ollie.
At the end I asked the room if they want the AI news segment to continue. Most people voted yes, so it stays.
Talk 2: Shipping a Paid AI Car Recommendation Platform with Claude
This talk was a real client project. The platform was built for a Korean fintech client. It is paid, and it is live in production on Azure. It was delivered in September.
The team was small and split between Australia and Korea. The speaker worked from Sydney with IT and marketing staff on the Korea side.
How it works
- The user types what they want in plain language.
- Claude Haiku extracts the intent using a 7-point rubric. Examples: fuel efficiency, brand origin, car type.
- The system scores cars from a Postgres database and returns the top 3. This scoring is deterministic.
- Claude Sonnet writes the description for each car.
Key design decisions
- The LLM extracts intent only. It does not choose the car. The same input gives the same result.
- An AI research agent fills data gaps. It searches the web for missing car data. But a human admin must approve the data before it is served.
- Guardrails. The user prompt is wrapped in a context layer before it reaches the Anthropic or OpenAI API.
Why not just use ChatGPT?
- The data is grounded in a real database.
- The output is structured.
- Invoicing and financing are in scope for phase 2.
How they test it
They keep a golden prompt set. Each prompt has an expected output. AI or a human scores the result. The pass threshold is about 95 to 98%.
They rerun the full set on every model change and every prompt change.
Timeline and tools
- Active work was about 3 months, between May and September.
- Without AI, the estimate was 28 weeks. AI coding saved about 16 weeks.
- Tools: Claude CLI, Claude Code, Claude API (Haiku and Sonnet), ComfyUI, ChatGPT, Gemini, ElevenLabs.
- The promotional video was made with ComfyUI animation plus an HTML overlay written by Claude.
The platform launched on September 28. In 2 days, 10 to 12 Korean media outlets covered it.
Lessons learned
- AI has no common sense without enough context.
- AI is overly agreeable. One example: when the user's budget was too low, it suggested secondhand cars instead of saying no.
- Say concurrency and throughput upfront. AI will not think about them on its own.
- AI changed how the team works, not just how fast. Testing, communication, and coding all changed.
Community Shout-Outs (Open Mic)
- Johannes (Amnells) — built an on-device text-to-speech app with no technical background. It has passed 1,000 installs.
- Richard (GlideStack) — an AI job application tracker. It matches your CV to a job description, writes cover letters, finds skill gaps, and shows all applications in one dashboard.
- Hongyi (Popwiz) — AI image and video production for fashion brands. Works with Shein and 10+ other brands. Has 12,000+ followers on Little Red Book. Finalist (6 of 109) in the Meitu competition, with possible investment.
- A real estate platform builder (Thailand) — looking for software ideas and collaborators.
- Shrini (FutureStream) — a product roadmap tool for teams that build AI agents. Targets agent builders in healthcare, legal, and real estate.
- A healthcare AI builder — works on clinical decision support and radiology image recognition. Looking to connect with GPs and doctors.
If you want to share next time, add your name to the open mic list when you arrive.
New This Month: Linky NFC Cards
Everyone got a Linky NFC card at registration. Tap it on a phone and it opens your LinkedIn profile. If you did not give your LinkedIn when you signed up, the card was blank.
The cards were free to use on the night. You can buy one to keep.
Get Involved
Want to speak at a future meetup, sponsor an event, or help shape the community? Get involved with 02Ship →
Speaker application form: https://forms.gle/CuhvaFTHKip9R1At9
Join the Next Meetup
The next meetup is at Haymarket HQ. Same time as before (6:00 to 8:30 PM).
It is free. Spots are limited.
See upcoming 02Ship meetups on Luma
See you there.