FUS-Meta: Offline-First AI & Optimization Framework
Run complex AI and optimization algorithms directly on edge devices—no cloud, no data leaks, full privacy.
- Automatically finds optimal neural architectures for MCU / FPGA / ASIC constraints
- Reduces manual tuning time by 30–70%
- Improves accuracy by 2–5% on noisy industrial data
- Optimizes power / latency / memory simultaneously
- Works fully offline for secure industrial environments
📄 Official Technical Report (CERN Zenodo):
https://zenodo.org/records/18343100?preview=1
The vision, architecture, and full benchmarks (98% accuracy, 47ms latency) are detailed in a citable technical report.
---## 🏭 Supported Industrial Hardware
✅ Arduino UNO Q (QRB2210/STM32U585) - 4GB RAM, dual-core ARM
• Cortex-A53: Full AutoML on-device training
• Cortex-M33F: Real-time execution on the factory floor
• Industrial endurance: -40°C to +85°C
🚀 Core Innovation Status: The novel architectural methods and optimization algorithms enabling the performance benchmarks below are the subject of a pending patent application. This repository contains public reference implementations and stable, obfuscated demonstrations for research, feedback, and non-commercial experimentation.
For inquiries regarding: Licensing of the core technology, commercial partnerships, or technical discussions under NDA, please contact via private message.
FUS-Meta is a framework and collection of tools for executing AI training and advanced optimization 100% locally on consumer or embedded hardware (Android phones, laptops, Raspberry Pi, ESP32). It is designed for use cases where data privacy, offline operation, or low-latency processing are critical.
🚀 COMING SOON: AutoFUS-MetaAI PRO – Enhanced version for powerful devices handling all CSV file types with results delivered directly to your phone. [Paid version launching soon]
"If it can run on a phone, it can run anywhere—privately."
AutoFUS-MetaAI = FUS-Meta + GravOpt + QuantumFUS + RLFUSMeta
Self-evolving AI that automatically designs neural networks for ASIC / FPGA / MCU / Edge devices
with power-aware, performance-aware, and memory-aware optimization.
The principles and quantitative results of the FUS-Meta approach have been formally documented. For a detailed analysis of the architectural vision, comparative benchmarks, and measured outcomes, see our technical preprint:
📄 FUS-Meta: A Vision and Benchmark for Fully On-Device, Privacy-Preserving Automated Machine Learning
🔗 View on arXiv.org – [LINK TO BE UPDATED UPON UPLOAD]
Key Validated Benchmarks (from Preprint):
| Metric | FUS-Meta / QuantumUSF | Traditional / Cloud-Based | Improvement |
|---|---|---|---|
| Object Recognition Accuracy | 98% | 85% | +13% pts |
| System Reaction Time | 47 ms | 150 ms | ~70% faster |
| Lane Stability (Autonomous Demo) | 94% | 78% | +16% pts |
| Braking Distance @ 50 km/h | 12.3 m | 16.1 m | 24% shorter |
| MAX-CUT (50k nodes) | 99.17% optimal | - | Solved in ~9s (CPU) |
| Project | Status | Description |
|---|---|---|
| FUS-Meta AutoML | 🟢 Public Beta | No-code AutoML for Android. Upload a CSV → get a trained PyTorch model, fully offline. |
| AutoFUS-MetaAI PRO | 🟡 Coming Soon | Enhanced version for powerful devices, supports all CSV formats, direct phone delivery. [Paid] |
| FUS-Meta Optimizer | 🔶 Stable Core | Quantum-inspired adaptive optimizer for large-scale problems (MAX-CUT, QUBO). |
| Edge AI Suite | 🔶 Research | Ultra-lightweight, self-adaptive neural networks for microcontrollers and edge devices. |
| GravOpt-MAXCUT | 🔶 Production | High-performance MAX-CUT heuristic for very large graphs (20k–100k nodes) on CPU. |
---AZURO CREATOR: AI that doesn’t just predict — it discovers new laws”
What does a scientist do?
Observes data → formulates a hypothesis → tests it → discovers a law.
And what do most AIs do today?
Learns a model → predicts → but doesn’t explain why.
AZURO CREATOR changes that.
🧠 This is the first system for truly automated scientific discovery that:
Generates human-understandable formulas (sigmoid, power, resonant, etc.)
Adapts the choice to the task:
→ Diagnostics: detects unexpected patterns (e.g. hidden phase transition)
→ Management: selects the most accurate models
Explains why a given hypothesis was chosen — through metrics for accuracy, novelty, and physical plausibility
Works completely locally — cloud-free, suitable for edge devices (even on a smartphone or ESP32!)
🔧 Applications:
Early diagnosis of defects in pumps, engines, hydraulic systems
Real-time anomaly detection (aviation, ships, power plants)
Automated science labs
Education: students see how laws are born
💡 Example:
We feed data with a hidden sigmoid transition → AZURO generates 9 hypotheses → selects “General + Sigmoid” → displays:
✨ “The system has a threshold behavior — probably a valve that opens at p1 ≈ 2.5”
🌍 The goal: To turn any device into an autonomous scientific agent, capable of discovering, explaining and learning.
🔗 Interested in a demonstration, collaboration or integration into your system?
👉 Write to me!
#AI #ScientificDiscovery #AutomatedScience #EdgeAI #IndustrialAI #MachineLearning #Innovation #BulgarianTech #AZURO #DigitalTransformation
For Android + Docker (local server):
- Download the beta package: FUS-Meta Beta v0.1
- Run the Docker container:
docker-compose up - Install the Android APK on your phone
- Connect your phone to your computer's local IP (same WiFi network)
- Upload a CSV file and train your model
🔜 PRO Version Preview: Soon you'll get enhanced processing for all CSV types with results directly on your phone.
pip install numpy numba
git clone https://github.com/Kretski/GravOptAdaptiveE
cd GravOptAdaptiveE
python example_maxcut.py
3. Edge AI on Microcontrollers
For ESP32, Jetson, or other embedded platforms, see the AzuroNanoOpt
🏥 Use Cases & Vision
Smart Cities & Autonomous Systems
Traffic Flow Optimization: Local, real-time analysis without streaming sensitive camera data.
Autonomous Vehicle Subsystems: High-performance, low-latency perception modules (see QuantumUSF benchmarks).
Healthcare & Medical Research
Train predictive models on sensitive patient data without leaving the hospital network.
Real-time analysis on medical IoT devices.
Industrial IoT & Automation
Solve logistics, scheduling, and resource-allocation problems on-premise.
Embedded predictive maintenance and anomaly detection on factory floor controllers.
Privacy-First AI
Financial data analysis, personal data mining, confidential business intelligence—zero data exposure.
💬 Community & Next Steps
Telegram Support Group
Join our Telegram group for beta testers and developers to get help, report bugs, and discuss edge AI:
FUS-Meta Beta Testers
We Are Actively Seeking
Beta Testers & Feedback: Especially from Healthcare, Industry 4.0, and Smart Infrastructure domains.
Research Collaboration: Academic or industrial research partners interested in edge AI and private ML.
Early Adopters for PRO: Organizations needing advanced, licensable technology for product embedding.
For all serious partnership, licensing, or collaboration inquiries, please initiate contact via private message on this platform.
📄 License
Core frameworks (Reference Implementations): MIT License
Commercial licenses & Core Technology: Available for enterprise embedding. Contact for terms.
AutoFUS-MetaAI PRO: Paid commercial license (launching soon).
👨💻 About the Author
Dimitar Kretski – Focusing on making advanced AI and optimization accessible, private, and deployable anywhere.
"The future of AI is not in bigger clouds, but in smarter, more private edges."