Handbook - Google DeepMind x Cactus Compute Global Hackathon
AI Tinkerers - San Francisco

Google DeepMind x Cactus Compute Global Hackathon

Handbook for the Google DeepMind x Cactus Compute Global Hackathon hackathon.

Final round winners have been announced. View Results

Handbook

Google DeepMind × Cactus Compute

Global Gemini Hackathon Handbook

Agentic Systems. Local-First. Edge + Cloud.


1. What This Hackathon Is

This is a five-city, one-day global build sprint organized by:

  • Google DeepMind
  • Cactus Compute
  • AI Tinkerers

Across multiple cities, curated rooms of engineers are building agentic systems that:

  • Run locally on-device using FunctionGemma
  • Execute via Cactus Compute runtime
  • Escalate to Gemini APIs in the cloud when needed

This is not a slide-deck hackathon.

It’s a room full of engineers shipping real systems — mobile, desktop, and hybrid — that push inference closer to the user without giving up frontier reasoning when necessary.

If you care about:

  • Local-first AI
  • Latency, privacy, offline capability
  • Agentic workflows that ship
  • Real trade-offs between edge + cloud

You’re in the right room.


2. The Technical Stack

🧠 Gemini (Cloud Reasoning Layer)

Role in this hackathon:

  • Complex reasoning
  • Long-context workflows
  • Heavy planning tasks
  • Fallback when local models aren’t enough

You are expected to:

  • Use Gemini APIs intentionally
  • Demonstrate clear escalation logic
  • Justify when and why cloud reasoning is required

🌵 Cactus Compute (Local Runtime)

Role in this hackathon:

  • Run AI locally on mobile devices and Macs
  • Enable low-latency agent execution
  • Reduce cloud dependency
  • Maintain offline capability

You should think about:

  • What must run locally?
  • What benefits from local execution?
  • Where does escalation make sense?

⚙️ FunctionGemma (On-Device Model)

Role in this hackathon:

  • Fast, lightweight on-device inference
  • Tool coordination locally
  • Immediate UX interactions

You are encouraged to:

  • Push FunctionGemma as far as it can reasonably go
  • Design agents that degrade gracefully
  • Treat cloud escalation as a decision — not a default

3. The Core Challenge

Build an agentic application that intelligently decides:

Where should this computation run — locally or in the cloud?

Your system should consider:

  • Latency
  • Privacy sensitivity
  • Offline constraints
  • Cost
  • Capability differences
  • Energy usage (optional but encouraged)

We are looking for architectural thinking, not just model usage.


4. Example Directions (Not Constraints)

These are prompts — not rails:

  • Offline-first AI assistant with cloud escalation
  • Local agents coordinating tools and workflows
  • Privacy-preserving AI apps that minimize data exfiltration
  • Hybrid inference pipelines across edge + cloud
  • New UX patterns unlocked by on-device intelligence
  • Context-aware switching between FunctionGemma and Gemini

If it ships and makes a strong architectural argument, it belongs.


5. Rules & Requirements

Team Structure

  • Teams of 1–4 builders
  • Solo builders welcome
  • Team formation deadline: 10:30 AM (local time)

Required Submission Components

Every team must submit:

  1. A working demo
  2. Architecture explanation
  3. Explicit use of:
  • FunctionGemma and/or
  • Gemini APIs
  1. Clear description of trade-offs

Projects without a functioning demo will not advance to finals.


6. Judging Criteria

We evaluate across five dimensions:

1️⃣ Architectural Clarity

  • Is the edge/cloud split intentional?
  • Is escalation logic explicit?
  • Are trade-offs defensible?

2️⃣ Technical Execution

  • Does it run?
  • Is the local inference real?
  • Is Gemini used meaningfully?

3️⃣ Agentic Design

  • Does the system plan?
  • Does it coordinate tools?
  • Is it more than a chat wrapper?

4️⃣ UX & Practicality

  • Does local execution improve the experience?
  • Would someone use this?

5️⃣ Ambition

  • Did you push the boundary?
  • Did you try something non-trivial?

We reward shipping, not theory.


7. Schedule (All Cities)

Time Activity
8:30 AM Doors open, breakfast & networking
9:00 AM Welcome + technical overview
10:00 AM Hacking begins
10:30 AM Team formation deadline
5:30 PM Project submissions due
6:00 PM Preliminary judging
7:00 PM Final demos
8:00 PM Winners announced
8:30 PM Event concludes

Local variations may apply by city.


8. Prizes

The following prizes are standardized across all cities:

🥇 1st Place — $5,000 in Google Cloud Platform (GCP) credits
🥈 2nd Place — $2,500 in Google Cloud Platform (GCP) credits
🥉 3rd Place — $1,000 in Google Cloud Platform (GCP) credits

Additional prizes may be announced on-site in each city.


9. What Makes This Different

This is an AI Tinkerers room.

That means:

  • Every attendee is curated
  • Most are actively shipping AI systems
  • Debugging happens live
  • Architecture debates are real
  • There is no “idea-stage networking”

We favor:

  • Demos over decks
  • Code over slides
  • Trade-offs over hype

If it doesn’t run, it doesn’t count.


10. Builder Expectations

You are expected to:

  • Respect other builders’ time
  • Ship something real
  • Be explicit about trade-offs
  • Share patterns you discover
  • Keep sponsor interaction technical

This is a technical room.


11. Media & Recording

The event may be photographed or recorded.

By attending, you consent to:

  • Event photography
  • Demo recordings
  • Short-form recap clips

If you need specific demo material removed for confidentiality, notify organizers immediately after your demo.


12. Data & Privacy

Default policy:

  • Aggregated reporting only
  • No personal attendee data shared without opt-in
  • GDPR / CCPA compliant handling

If you escalate data to cloud inference, explain how you handle:

  • User consent
  • Sensitive inputs
  • Storage decisions

Architectural ethics matter.


Final Reminder

This is not a hype event.

It’s a builder room.

Push the boundary.
Ship something real.
Explain your trade-offs.

Code talks. Hype walks.