Skip to content

GarrickPinon/SwissblAIz

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 

Repository files navigation

🌵 SwissblAIz V3 — Hybrid Voice Assistant

Built for the Google DeepMind × Cactus Compute Hackathon

Google DeepMind Cactus Compute FunctionGemma Gemini Flash

On-device voice assistant powered by FunctionGemma-270M + Gemini Flash cloud fallback


🎙️ Voice-Activated Multi-Tool Agentic Demo

Demo Video

Try it live: Open demo/index.html in Edge or Chrome.

Features:

  • 🎤 Voice-first — Tap the orb, speak naturally
  • On-device inference — 50-80ms via FunctionGemma-270M
  • ☁️ Cloud escalation — Auto-routes hard queries to Gemini Flash
  • 🇦🇺 Australian male voice — TTS responses via Microsoft Neural voices
  • 📊 Real-time pipeline — See STT → Classify → Route → Infer → TTS live

Architecture

User Query (Voice / Text)
    │
    ▼
┌─────────────────────────┐
│  COMPLEXITY ROUTER      │  ◄ Deterministic, <1ms
│  EASY | MEDIUM | HARD   │    No LLM call
└────────┬────────────────┘
         │
         ▼
┌─────────────────────────┐
│  Cactus SDK + Gemma     │  ◄ Unified hybrid path
│  threshold → routing    │    SDK handles escalation
└────────┬────────────────┘
         │
    ┌────┴────┐
    ▼         ▼
┌────────┐ ┌──────────┐
│ ⚡ Local│ │ ☁️ Cloud  │
│ 50-80ms│ │ 200-500ms│
│ GemMA  │ │ Gemini   │
└───┬────┘ └────┬─────┘
    │            │
    ▼            ▼
┌─────────────────────────┐
│  TRN VALIDATOR          │  ◄ Schema + type check
│  + Postprocessor        │    F1 normalization
└─────────────────────────┘

Key Innovations

  1. Cactus SDK Unified Path — Single cactus_complete call with native hybrid routing
  2. Cloud Fallback — SDK-managed escalation to Gemini Flash for complex queries
  3. Deterministic Complexity Router — No LLM needed for routing, zero latency overhead
  4. TRN Validator — Validates every tool call against its JSON schema
  5. Voice-First UI — Animated orb, action cards, pipeline telemetry

Running

# Setup (one-time)
git clone https://github.com/cactus-compute/cactus
cd cactus && source ./setup && cd ..
cactus build --python
cactus download google/functiongemma-270m-it --reconvert
pip install google-genai requests
export GEMINI_API_KEY="your-key"

# Run voice demo
open demo/index.html

# Submit to leaderboard
python submit.py --team "SwissblAIz" --location "Online"

Team

SwissblAIz 🌵

Built with ❤️ for the Google DeepMind × Cactus Compute Hackathon

Powered by FunctionGemma-270M (on-device) + Gemini Flash (cloud)

About

SwissblAIz is a model‑agnostic agentic controller built for the Google DeepMind × Cactus Compute Global Hackathon. Orchestrates on‑device Gemma/LFM models with deterministic routing, reflection‑aware tool use, and a hybrid‑compute execution layer. 9th Place. The system focuses on reliability over flash: clean tool‑call F1, stable reasoning loops, e

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors