A Sleek, Stable, and High-Performance Local AGI Platform for Windows & Android
Cyborg AGI was built specifically around the Gemma 4 Good Hackathon challenge: use the power of open, locally-runnable AI to solve real-world problems in health, education, and global resilience — especially in low-connectivity and privacy-sensitive environments.
| Hackathon Theme | How Cyborg Addresses It | Status |
|---|---|---|
| 🏥 Health | Gemma 4 Health pipeline for local clinical reasoning & X-ray analysis — no patient data leaves the device | ✅ Completed |
| 🎓 Education | Adaptive Tutor with OCR-graded homework, dynamic quizzes, and culturally localized content | ✅ Completed |
| 🌍 Global Resilience | Offline-capable architecture + World Monitor for geopolitical & disaster intelligence | ✅ Completed |
| 🏢 Enterprise & Economy | Zero-Employee Company (Company OS) via Paperclip Workspace for local autonomous tasking | ✅ Completed |
| 📞 Agentic Voice & Outbound | Outbound Twilio calling for automated lead generation, customer support, Whisper/Kokoro STT/TTS | ✅ Completed |
| 🔒 Privacy-First | Fully local inference — no cloud dependency, FHIR-compatible data guardrails | ✅ Completed |
| ⚙️ Edge & Constrained Environments | Runs on Windows workstations, Android, Raspberry Pi, and NVIDIA Jetson | ✅ Completed |
Why Gemma 4? Open Apache 2.0 weights, strong multimodal performance on local hardware, and safety-tuned by design. The exceptional agentic intelligence of Gemma 4 enables complex reasoning, planning, and tool parsing at the edge, making the entire on-device AGI workflow highly optimized, responsive, and efficient.
Experience a live interactive sample of Cyborg AGI directly in your browser:
- 🌐 Live Space Demo: https://huggingface.co/spaces/cyborg2005/cyboorg
Cyborg AGI's ultimate vision is to create a seamless, decentralized, private network connecting all your personal devices (PCs, laptops, and Android phones) into an air-gapped, zero-cost private grid, expanding into a partially sentient, self-healing system:
- Seamless Device Interconnectivity: All your devices remain connected securely in real-time, allowing you to access the combined power of your local ecosystem from anywhere.
- Natural Language File Retrieval: Fetch any file from any of your devices simply by describing what you need in natural language. For instance, even if a file resides on your desktop PC at home, you can query it on your phone while you are outside, and the system will locate and stream it directly to your phone.
- Remote Task Orchestration & Management: Delegate active workloads and assign tasks directly from your Android device. Setup new coding environments, bootstrap whole programming projects, or trigger custom server tasks on your remote PC while you are on the move.
- Self-Generating Code & Tool Creation (Sentient Agents): If you ask the system to perform a new action that it doesn't currently support (e.g., "open WhatsApp") and the code or tool to do so is missing, the system will automatically generate the necessary code as a LangGraph or Deep Agent tool at runtime. It compiles and injects it into its own source code to execute the task on the fly—rendering the platform partially sentient and capable of infinite self-growth!
- 100% Free, Private, & Local: No subscriptions, no telemetry, and no third-party servers. Your files, local LLMs, and operations are kept strictly local, free, and private forever.
Cyborg AGI is a full-stack, locally-hosted Autonomous General Intelligence OS. It unifies a Flutter-powered dashboard, a FastAPI intelligence backend, and a personal Knowledge Vault into a single coherent system — all running on your own hardware.
Think of it as your own private Jarvis: it reasons, remembers, monitors the world, teaches, and assists with clinical questions — without sending a single byte to a third-party cloud.
- 🧠 General-purpose chat powered by Llama/Qwen local models
- 🏥 Medical reasoning via isolated Gemma 4 Health with SigLIP vision
- 🎓 Adaptive education with homework grading and personalized quizzes
- 🌎 World Monitor for live geopolitical and disaster intelligence
- 🕸️ Knowledge Graph (Mirofish engine) for semantic memory visualization
- 🎙️ Jarvis Voice with local Whisper STT and Kokoro TTS
- 📂 Integrated File Explorer connected to your Obsidian-compatible knowledge vault
- 🔄 Autonomous GitHub Sync for continuous self-improvement pipelines
Cyborg is organized into three tightly integrated layers:
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ Frontend Shell (Flutter 3.x) │
│ Dashboard · World Monitor · Mirofish Graph · Paperclip UI · Voice Agent Workspace │
└────────────────────────┬────────────────────────────────────────────────────────────┘
│ REST / WebSocket
┌────────────────────────▼────────────────────────────────────────────────────────────┐
│ Backend Intelligence (FastAPI + Python) │
│ General Chat │ Gemma 4 Health │ Adaptive Tutor │ Company OS │ Twilio Voice Caller │
└────────────────────────┬────────────────────────────────────────────────────────────┘
│
┌────────────────────────▼────────────────────────────────────────────────────────────┐
│ Knowledge Vault (ACE/Obsidian Structure) │
│ .md Semantic Nodes · Wikilink Graph · FHIR EHR · Lead Campaigns │
└─────────────────────────────────────────────────────────────────────────────────────┘
graph TD
subgraph "Frontend (Flutter Shell)"
UI["📱 Dashboard / World Monitor"]
KG["🕸️ Mirofish Knowledge Graph"]
GS["⚙️ Global Settings & Dark Mode"]
MF["🐠 Mirofish Workbench"]
VX["🎙️ Jarvis Voice Assistant"]
PC["📎 Paperclip Workspace & Theme"]
VAG["📞 Outbound Voice Calling Panel"]
end
subgraph "Backend Intelligence (FastAPI)"
API["🐍 API Gateway"]
LLM["🧠 General Chat (Llama / Qwen)"]
MED["🏥 Gemma 4 Health — Isolated Pipeline"]
EDU["🎓 Adaptive Tutor (Gemma 4 Weights)"]
KGE["🌌 Graph Physics Engine (Leiden)"]
WM["🌎 World Monitor — GDELT / USGS Feed"]
COMP["🏢 Company OS — Paperclip Workflow"]
TWIL["📞 Twilio Calling Engine (edge-tts / STT)"]
end
subgraph "Knowledge Vault (ACE Structure)"
VAULT["🗄️ Obsidian-Compatible Vault"]
GRAPH["🔗 .md Wikilink Semantic Graph"]
ACE["📂 Atlas / Calendar / Efforts"]
FHIR["🔐 FHIR-Compatible EHR Store"]
LEADS["🗃️ Leads Database & Call Logs"]
end
UI <--> API
VX <--> API
KG <--> KGE
PC <--> API
VAG <--> API
API <--> LLM
API <--> MED
API <--> EDU
API <--> WM
API <--> COMP
API <--> TWIL
LLM --> VAULT
KGE --> GRAPH
MED --> FHIR
EDU --> VAULT
COMP --> VAULT
TWIL --> LEADS
VAULT --> ACE
Important
Hackathon Focus Area: Privacy-preserving medical AI for environments where patient data cannot leave the device.
Cyborg's Health Track utilizes the global, highly optimized LLMService to perform fast, on-device clinical reasoning and vision tasks using Gemma 4.
- Unified Vision Projector Architecture: Routes all chest X-ray image encodings directly through the central GPU-accelerated vision projector, completely eliminating separate, redundant vision model loads and preventing VRAM thrashing.
- Smart RAG Context Guardrails: Features a robust, real-time context cleaner inside health_edu.py that screens out unrelated machine learning or system lecture notes from the patient history context, ensuring 100% focused clinical assessments.
- Local Clinical Reasoning: Gemma 4 reasoning runs entirely on-device — suitable for clinics with no internet access or strict HIPAA/DPDP compliance requirements.
- Zero-Exfiltration Architecture: Patient records, images, and diagnostic outputs remain in the local encrypted vault.
lib/health/gemma4/ (formerly medgemma)
├── inference.py # Optimized async multimodal X-ray pipeline (centralized LLMService)
├── prompts.py # Medical prompt templates & mandatory disclaimer injection
└── ehr_functions.py # FHIR function-calling with hardcoded safety guardrails
- User uploads a chest X-ray or describes symptoms via the Flutter dashboard.
- The endpoint queries local vault medical notes (automatically screening out unrelated technical lecture slides).
- The image is routed async to the global
LLMServicemodel, passing the processed chest scan to the vision projector. - Output includes a structured differential, mandatory disclaimers, and optionally a FHIR-compatible EHR record.
- All data stays in the local vault under the
Atlas/Health/ACE directory.
# Start the Health demo server (Port 7860)
python assets/demos/health_demo.py
# Visit: http://localhost:7860Hackathon Focus Area: Personalized, offline-capable learning tools for students in rural or under-resourced classrooms.
The Education Track leverages the centralized Gemma 4 weights in LLMService to power an adaptive tutor that evaluates work and teaches without requiring any external cloud APIs.
- Multimodal Homework Grading: Students photograph handwritten work; the global
LLMServicereads the image directly via the shared vision projector, evaluates correct solutions, and outputs grades with 0.95+ confidence. - Adaptive Quiz Generation: After grading, the tutor identifies specific knowledge gaps and generates targeted follow-up questions — no two sessions are the same.
- Culturally Localized Content: Teaching examples and problem contexts are adjusted for India, US, and SE Asia, making explanations more relatable.
- Progress Analytics Dashboard: Learning trajectories are tracked locally, helping teachers identify students who need intervention.
lib/education/adaptive_tutor/
├── grader.py # Optimized async multimodal grader (OCR-free direct vision projector)
├── quiz_generator.py # Dynamic quiz creation targeting identified gaps
└── progress_tracker.py # Learning analytics and path optimization
- Student submits a photo of handwritten homework via the app.
- OCR extracts the text; Gemma 4 evaluates correctness against the subject rubric.
- The grader identifies weak concepts and passes them to the quiz generator.
- The tutor generates a personalized set of follow-up questions with culturally relevant examples.
- All progress data is stored locally in the knowledge vault for longitudinal tracking.
# Start the Education demo server (Port 7861)
python assets/demos/education_demo.py
# Visit: http://localhost:7861Tip
Supercharging Local Operations: Run an entire enterprise of research and task execution completely autonomously with local intelligence.
Cyborg AGI integrates Paperclip, an advanced enterprise workspace and theme that hosts the Zero-Employee Company (Company OS). This feature harnesses the reasoning depth of Gemma 4 to execute highly professional business, research, and coding tasks completely autonomously.
- Autonomous Research Agents: Researchers query topics, compile literature reviews, and perform deep-dive web or local knowledge assessments.
- Task Orchestration Pipelines: The local LLM functions as a professional project manager—researching topics, creating plans, writing scripts, and managing code workflows.
- High-Fidelity Local Execution: Performs complex business tasks on-device without subscription fees, corporate telemetry, or cloud dependencies.
Hackathon Focus Area: Climate, disaster response, and geopolitical awareness in one real-time intelligence feed.
The World Monitor provides an AI-synthesized, live situational awareness layer — directly relevant to the hackathon's Global Resilience track.
- 🛰️ Real-Time Map: Conflict zones, natural disasters, and instability hotspots visualized on a high-performance vector map.
- 📊 AI-Driven Daily Briefing: Automated synthesis of GDELT news events, USGS earthquake data, and market risk indicators into concise actionable summaries.
- 📉 Instability Scoring: Proprietary country-level stability index combining multi-source signals, updated every 5 minutes.
- 📡 Low-Latency Feeds: News tickers and strategic overviews streamed directly into the AGI context for downstream reasoning.
Cyborg's memory system is not a flat vector database — it is a living, force-directed semantic graph inspired by Obsidian and Mirofish.
- Force-Directed Physics Engine: Nodes (markdown files) and edges (wikilinks) form natural knowledge clusters via repulsion/attraction simulation.
- Interactive Exploration: Drag to pin nodes, click for a floating details panel, adjust Repel Force and Center Gravity in real time.
- Leiden Community Detection: Automatically color-codes related knowledge clusters.
- Semantic Triplet Extraction: During document ingestion, Cyborg extracts subject-predicate-object triplets and creates explicit
[[Wikilinks]], populating the graph organically. - Zero-Deadlock State Management: Powered by an optimized Riverpod engine resolving all previous UI race conditions.
Cyborg AGI features a fully local voice interface combined with an outbound telephony engine to solve real-world communication and scheduling problems at zero cost:
- Lead Generation & Outbound Campaigns: Run bulk outbound call campaigns to generate sales, qualify prospective clients, and collect contact details using your custom Twilio telephone number.
- Automated Customer Support: Configures automated regional support desks that handle incoming queries, log customer pain points, and classify intents locally.
- Human-like Telephony Speech: Uses advanced, ultra-realistic neural voice synthesis (
edge-tts) with instant fallback to offline local engines (pyttsx3). - Lead Qualification Pipeline: The local Gemma 4 parses the user's intent in real-time, marks leads as
interested,not_interested, orcallback, and updates the persistent JSON database with CSV export capabilities.
- Hands-Free Operations: Fully local Whisper STT for accurate, accent-robust speech recognition with zero cloud API dependency.
- High-Fidelity Synthesis: Local Kokoro ONNX for gorgeous, emotionally resonant voice output.
- Full-Duplex Interruption: Interrupt the AI mid-sentence by simply talking or typing—creating a natural, lifelike flow.
All intelligence is anchored in an Obsidian-compatible vault organized by the ACE Synthesis Framework:
vault/
├── Atlas/ # Maps of topics, people, places — reference knowledge
│ └── Health/ # Gemma 4 Health outputs & EHR records
├── Calendar/ # Time-anchored notes and World Monitor briefings
└── Efforts/ # Active projects and adaptive learning paths
Every model interaction (health, education, general chat) writes back to this vault, creating a compounding personal knowledge base over time.
Cyborg is designed to run where cloud AI cannot:
| Target Device | Notes |
|---|---|
| Windows 10/11 Workstation | Primary platform, CUDA-optimized |
| Android Phone/Tablet | APK via Flutter; edge model variants |
| Raspberry Pi 5 | ARM-compatible, quantized models |
| NVIDIA Jetson Nano/Orin | Full GPU inference at the edge |
# Deploy to any edge device
bash scripts/deploy_gemma4_edge.sh- Python 3.10+
- Flutter 3.x
- CUDA 12+ (optional, for GPU acceleration)
- Docker (optional, for containerized deployment)
cd assets/backend
python setup_env.pyflutter pub get
flutter run -d windows # Windows
flutter run -d android # Android# Install Firebase CLI
npm install -g firebase-tools
# Configure FlutterFire
dart pub global run flutterfire_cli:flutterfire configure
# Place google-services.json in android/app/ then:
python sync_firebase.py
flutter pub get
flutter run -d windowscp .env.example .env # Edit with your settings
docker-compose up --build # Starts backend on port 8765API Endpoints once running:
http://localhost:8765/api/v1/health— Health checkhttp://localhost:8765/api/docs— Swagger UIhttp://localhost:8765/api/redoc— ReDoc
# .env (copy from .env.example)
# LLM Backend
DEFAULT_MODEL=qwen2.5-coder-14b
CONTEXT_LENGTH=4096
N_GPU_LAYERS=-1 # -1 = all layers on GPU
# Embeddings
EMBEDDING_MODEL=all-MiniLM-L6-v2
EMBEDDING_DEVICE=cpu
# Features
ENABLE_VOICE=true
ENABLE_WORLD_MONITOR=true
OFFLINE_MODE=false # Set true for fully air-gapped deployment| Real-World Pain Point | Cyborg Local Solution | Social/Economic Value |
|---|---|---|
| High SaaS Subscription Costs | Zero subscription fees. Runs fully local AGI workflows (research, analysis, grading) on your desktop. | Saves $500+/year for indie developers, students, and small enterprises. |
| Expensive Cold Outreach / Staffing | Agentic Voice Calling handles support and outbound lead generation automatically using Twilio and local LLMs. | Empowers local solopreneurs to scale outbound campaigns with zero staffing overhead. |
| Strict Patient/Client Privacy Rules | Gemma 4 Health differential diagnosing and chest scan analysis runs strictly local in your air-gapped vault. | Eliminates telemetry risk, fully compliant with strict HIPAA and GDPR standards. |
| Information Overload & Scattered Files | Mirofish Knowledge Graph dynamic Leiden-clustering visualizes notes, calendar activities, and links. | Solves the ADHD/knowledge-worker indexing problem by offering organic, compounding visual memory. |
| Isolated Rural Clinics & Classrooms | Gemma 4 Health (X-ray analysis) and Adaptive Tutor (homework scanner) require exactly 0% internet. | Closes the digital divide, providing cutting-edge educational & clinical help in remote global locations. |
Cyborg AGI is designed to solve high-stakes, real-world problems where cloud dependency or corporate AI isn't viable:
- Empowering Remote Classrooms: In rural villages or schools without reliable internet connections, students and teachers cannot access ChatGPT or Claude. With the Cyborg Education Window, they can photograph handwritten assignments. The system grades the submission on-device, highlights learning gaps, and generates personalized quizzes instantly.
- Culturally Relevant Learning: By adapting quiz topics to regional demographics (e.g., India vs. US vs. SE Asia), it ensures local students learn using concepts, currencies, and scenarios they actually understand.
- Offline Clinical Aid: In rural or under-resourced diagnostic centers, prompt clinical reasoning can be a matter of life and death. The Cyborg Health Window analyzes chest X-rays locally on-device. It scans patient records, differential diagnoses, and suggests treatments without exposing private patient data to external servers, providing immediate triage support.
- Context-Aware Medical History: Seamlessly integrates personal history from the local RAG vault while screening out technical noise (such as unrelated ML lecture slides), producing pristine medical evaluations.
- Compact Local Benchmarks: All of our development, validation, and benchmarking were performed on 2B and 4B models due to local hardware and VRAM constraints.
- Outstanding Efficiency: Despite these constraints, the optimized Gemma models performed exceptionally well—beating many general cloud endpoints in task-specific accuracy and response latency.
- Future Scaling Roadmap: We plan to scale Cyborg AGI to support larger parameter models in future iterations, making local, private AGI accessible and useful for thousands of people worldwide.
| Metric | Value | Status |
|---|---|---|
| Inference Speed | 60+ tokens/sec (RTX 3070+) |
⚡ Premium |
| Gemma 4 Health VRAM | ~6 GB (4-bit quantized) |
🔋 Optimized |
| Paperclip Company OS | Multi-agent research loop (local only) |
💼 Free Operations |
| Voice Assistant Latency | ~120ms Whisper STT / ~200ms TwiML loop |
🔊 Low Latency |
| Knowledge Graph | Tested to 10,000+ nodes |
🕸️ Leiden community active |
| Cold Start Time | < 8 seconds (Windows native) |
🚀 High speed |
| Offline Capability | Full (all features 100% air-gapped) |
🔒 Air-gapped |
- Container runs as non-root user (
cyborg, UID 1000) - All sensitive config via environment variables — never baked into images
- Gemma 4 Health pipeline is fully isolated from general LLM routing
- FHIR EHR store uses hardcoded ethical guardrails that cannot be overridden via prompt
- No telemetry, no cloud callbacks, no third-party data sharing
- Gemma 4 E4B edge model integration for Android offline-first deployment
- Climate track: local energy optimization advisor (Global Resilience)
- Federated Knowledge Vault sync across local network nodes
- Expanded FHIR function-calling for medication management
- Multilingual TTS/STT for regional language support (Hindi, Tamil, Bahasa)
| Resource | Link | Description |
|---|---|---|
| 🏗️ Architecture Deep-Dive | GEMMA4_QUICKSTART.md |
Full local Gemma configuration guide |
| 💼 Company OS Design | paperclip_integration_plan.md |
Zero-Employee Company integration roadmap |
| 🎙️ Voice Integration Summary | voice_agent_summary.md |
Telephony Twilio Voice Cold-Calling guide |
| 🔌 API Reference | http://localhost:8765/api/docs |
Swagger interactive API gateway UI |
| 🏆 Hackathon Submission | Kaggle Competition Page | Hackathon details and deliverables |
| 🧠 Gemma 4 Models | Kaggle Models · Hugging Face | Official Gemma model source paths |
Built with ❤️ using Gemma 4 for the Gemma 4 Good Hackathon · Kaggle × Google DeepMind · 2026
"Stable Intelligence. Autonomous Growth. AI for Everyone, Everywhere."
