BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//claudecommunity.com.au//Events//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:Claude Code Community Australia Events
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-PUBLISHED-TTL:PT1H
BEGIN:VEVENT
UID:cmq7fergx0000psp7lkdm39ms@claudecommunity.com.au
DTSTAMP:20261115T221217Z
LAST-MODIFIED:20261115T221217Z
DTSTART:20261118T063000Z
DTEND:20261118T083000Z
SUMMARY:Sydney | Claude for Developers
DESCRIPTION:17:30 – 18:15 | Doors open\, food\, drinks and networking\n\n
 18:15 – 18:25 | Welcome and intro\n\n18:25 – 18:45 | From Vibe Coder t
 o Builder: Shipping an AI App with Claude Code in 5 Days — Lingyun Zhao\
 n\nVibe coding gets you a demo. Getting to something you can trust takes m
 ore.\n\nI'm building DrainWise solo with Claude Code for the City of Parra
 matta hackathon. It's an AI tool that predicts which stormwater drains wil
 l block next\, so crews can clean the riskiest ones first.\n\nIn this talk
 \, I'll show how I actually build with it:\n\n	•	Setup and workflow: pro
 ject brief\, context handoff\, keeping the agent on track\n	•	Agents acr
 oss the stack: frontend\, backend\, auth and database\n	•	What broke\, a
 nd what I'd do differently\n	•	Next step: RAG and an AI system that lear
 ns from feedback\n\nTakeaway: AI makes building fast. The workflow around 
 it makes it shippable.\n\nI'm a Sydney-based software engineer and indie f
 ounder with a PhD in Computer Vision from Queen Mary University of London 
 and around seven years of full-stack experience. I'm building EvidMatch\, 
 an AI-powered talent marketplace\, and DrainWise\, a hackathon project tha
 t predicts blocked stormwater drains\, both with Claude Code. I'm interest
 ed in AI-assisted development\, practical AI adoption\, and turning techni
 cal ideas into real products.\n\n18:45 – 19:05 | The Benchmark Was Wrong
 : What Claude Found When We Gave It Context Instead of SQL — Pete Chapma
 n\n\nEnterprise data is the hill that AI agents still die on. Using BEAVER
 \, MIT’s university data warehouse benchmark\, agents writing SQL answer
 ed just 26% of questions correctly. Two-thirds of their answers were confi
 dently wrong.\n\nWe ran the same questions through Claude Opus 5.5\, GPT-6
 .1 Sol and Grok 4.7\, this time connected over MCP to an Amorfs knowledge 
 pack: a structured\, typed projection of the same warehouse\, with joins p
 re-resolved and business meaning made explicit. Accuracy rose to 97%.\n\nT
 hen something amazing happened. Asked simply to review the benchmark\, Cla
 ude answered every question in three minutes and then flagged discrepancie
 s in the official answers. That review led to 18 proposed corrections\, no
 w sent to the benchmark’s authors.\n\nThis talk covers how the pack is b
 uilt\, what agents need from context that schemas don’t give them\, and 
 the design patterns that bring context into your own MCP servers.\n\nPete 
 Chapman is founder and CEO of First Cognition\, builder of Amorfs\, a cont
 ext layer for AI agents. He has a background in cognitive science and ente
 rprise architecture\, and is based in Sydney and Ho Chi Minh City.\n\n19:0
 5 – 19:25 | One founder\, an agent team: shipping a full business platfo
 rm (web\, iOS\, Android\, payments\, door access) to production — Prasha
 n Narayanasamy\n\nMovena is an operating system for gyms\, from a single l
 ocation to multi-site groups: timetable and bookings across locations\, me
 mberships and Stripe payments\, door access\, a gym website with an embedd
 able booking widget\, member apps on iOS and Android(HealthKit and HealthC
 onnect)\, and a connector in the ChatGPT app directory and meta/google cam
 paign connector for leads. It is built and run by one person.\n\nThis talk
  is about the operating model that makes that possible\, not about prompts
 .\n\nWhat I'll cover\nLanes\, not one big chat. Separate Claude Code sessi
 ons own backend\, review and infrastructure\, each in its own git worktree
 \, with rules about what each may touch\, deploy or message.\nThe founder 
 locks behaviour\; agents don't. Every defect is walked through and its beh
 aviour locked before any code is written. Codex implements from a paste-re
 ady brief\; Claude reviews.\nReviewers are not the final authority. How "i
 ndividually reasonable" review suggestions quietly grew a feature from two
  levels to six\, and the written rule that stopped it.\nTwo reviewers beat
  one. A real example from this week: my review passed a widget change\; an
  independent review lane proved with a test that the new route was missing
  from a proxy allowlist and every Book tap would fail. It never reached a 
 customer.\nGuardrails that actually hold: deploys only on an explicit go i
 n chat\, contracts for released mobile APIs\, one defect tracker of record
 \, help pages shipped with every feature\, and money changes tested end to
  end against the Stripe sandbox.\n\nWhat you'll walk away with: a repeatab
 le loop (walkthrough → lock → brief → independent review → UAT →
  production) and the specific rules I wish I had written down on day one.\
 n\nhttps://movena.com.au\n\nPrashan is the founder of Movena\, an operatin
 g system for gyms from a single location to multi-site groups\, spanning w
 eb\, iOS and Android apps\, payments and door access. He builds and runs i
 t solo with a team of AI coding agents and ships to production several tim
 es a week.\n\n19:25 – 19:45 | Stop benchmarking models\, benchmark your 
 harness — Jack Rudenko\n\nEveryone benchmarks models. Almost nobody benc
 hmarks the harness around the model\, and that is where most of the varian
 ce in real work lives.\n\nWe built an internal eval rig to test our own Cl
 aude Code setup against stock Claude Code and other harnesses on real repo
 s. Most of the config folklore did not survive contact with the numbers.\n
 \nWhat I'll cover:\n\nHow to build a repeatable eval for a coding harness:
  task sets from real repos\, graders\, cost and wall clock as first class 
 metrics\, repeat runs so one lucky pass does not fool you\nWhat actually m
 oves the needle (context discipline\, tool surface\, loop shape) vs what d
 oes nothing or makes things worse (piling on MCP servers\, a giant CLAUDE.
 md\, subagents for everything)\nModel vs harness: a frontier model bare ag
 ainst a cheaper model in a good harness\, and what the price gap actually 
 buys you\nLive run: stock setup vs tuned setup\, read the numbers on stage
 \n\nWalk away with: a method to bench your own setup this week\, the metri
 cs worth tracking\, and a shortlist of config changes that measurably pay 
 off.\n\nFormat: 20-25 min with the live demo\, or a 10-min lightning cut w
 ithout it.\n\nCo-founder and Chief AI Officer at MadAppGang in Sydney. Twe
 nty five years of custom software delivery\, now mostly building agentic h
 arnesses\, evals and tooling for teams running AI in production.\n\n19:45 
 – 20:00 | Speaker panel Q&A\n\nAll four speakers take questions from the
  audience.\n\n20:00 – 20:10 | Community announcements and close\n\nhttps
 ://claudecommunity.com.au/events/sydney-claude-for-developers-november-202
 6\nTickets: https://luma.com/claude-w21b
URL:https://claudecommunity.com.au/events/sydney-claude-for-developers-nove
 mber-2026
LOCATION:AWS Sydney\, Sydney
END:VEVENT
END:VCALENDAR
