GLOBAL AI CONSTRUCT- BRISBANE

# Global AI Construct - Brisbane

Stop watching. Start building.
Global AI Construct is a hands-on event format where participants hack, build, and experiment together. No spectators—just tools, peers, and something real to demonstrate by the end of the event.

Agenda:

1. 5:15 PM- Registration, snacks and networking
2. 5:45PM - Welcome and setting the agenda, speaker and lab introduction
3. 6:00 PM- 1.25 hours of Agentic BUILD with Jernej Kavka
4. 7:30 PM - 1 hour of Agentic BUILD with Rohit BANSAL
5. 8:30 PM- Closing notes and Q/A

This is what we will build on the day of the event :
LAB 1

• Small Models, Sharp Jobs: Build a Local Agent That Shows Its Evidence
• In plain English, attendees build a .NET agent that reads a fictional incident pack using a safe local tool. A small open-weight model running on the laptop proposes claims and quotes. Ordinary C# checks the evidence, dates, numbers and supported event phrases, keeps uncertainty visible, and renders a cited Markdown brief. Attendees then break a fact on purpose and watch their own verification rule catch it.
• Set up instructions will be given a week prior to the event so that Builders can come prepared with the pre-requisites like LM Studio, VS Code, SLMs and more.
• We will update the instructions here for the lab preparation.
• LAB INSTRUCTOR AND SPEAKER : Jernej Kavka, Microsoft AI MVP, Senior AI SME (SSW) (Jernej Kavka | LinkedIn)
• PRE REQUISITES for LAB 1 :
• Before the lab — please download on home/office Wi-Fi, not at the venue.
  1. .NET 10 SDK — https://dotnet.microsoft.com/download/dotnet/10.0 (dotnet --version should print 10.x)
  2. A local LLM runtime — LM Studio is simplest (lmstudio.ai). Ollama also works (ollama.com).
  3. One model, downloaded in advance.
  Recommended for everyone — Nemotron 3 Nano 4B (~2.8 GB). Smallest, fastest, tested end-to-end for this lab and can work without GPU.
  Got a MacBook with Apple Silicon, or a machine with plenty of VRAM? These suit that hardware better:
  - 16 GB Apple Silicon, or 8+ GB VRAM → gemma4:12b-it-qat (~7 GB), or Bonsai 27B (~9 GB, experimental)
  - 32 GB+ Apple Silicon, or 24+ GB VRAM → gemma4:26b (~16 GB), or Qwen3.8 27B (~16 GB)
  Partial GPU offload is fine — you don't need the whole model in VRAM.
  Bigger is not better here. The 4B matches all of them on this lab's checks. That's rather the point of the session.

LAB 2

• RAG IN SNOWFLAKE— quietly keeping watch through the day. Proactive inbox + calendar chief-of-staff
  The heartbeat wakes every 15 minutes, scans email and calendar, and messages you on Telegram/WhatsApp only when something needs you: "This client email needs a reply before your 2pm — draft ready, want me to send?" It reasons about what's actually urgent instead of buzzing on everything. Plays directly to the proactive heartbeat + messaging surface.
• PRE REQUISITES FOR THE LAB 2: https://sites.google.com/view/rag-chatbot-on-snowflake/home
• LAB 2 instructor : Rohit Bansal, Rohit Bansal | LinkedIn