Nuclear Intelligence: Open-Source AI for Nuclear Research

Nuclear Intelligence is an open GitHub project providing an autonomous pipeline for nuclear energy research. The system converts verified Q&A pairs into NES tokens recorded on blockchain using SHA-3 Proof-of-Work. Designed for scientific organizations and developers working with peaceful nuclear technologies.

Key takeaways

• Open GitHub repository with 300+ commits, last updated June 2024
• Automated research cycles every 25 minutes via GitHub Actions
• Supports 6 programming languages including Python and JavaScript
• Three-tier data validation: ≥70% accuracy, ≥60% novelty, ≥60% usefulness
• Tokenized outputs: 1 NES per successful research cycle
• Public data mirror on HuggingFace updates every 2 hours

Architecture overview

The core system comprises three interconnected modules. The LLM Engine processes queries with support for 7 AI providers including Groq and Gemini.

Core components

• RAG FAISS — Vector database for scientific material storage/retrieval
• Knowledge Graph — Semantic network of concept relationships
• Virtual Ledger — Blockchain for NES token transactions

Additional modules

• Web Search — Integration with arXiv and scientific databases
• Enhanced API — FastAPI interface for programmatic access
• Health Check — System monitoring with 9 diagnostic tests

Technical workflow

The system employs multi-stage data processing. Layer 1 analyzes queries via RAG (Retrieval-Augmented Generation), Layer 2 validates against Knowledge Graph, Layer 3 generates tokenized outputs. An optimized pipeline processes up to 50 scientific papers per cycle.

Security and ethics

Two-tier content filtering: Pre-filter blocks weapons-related queries, post-filter verifies outputs against IAEA standards.

Restriction type	Blocked topics
Nuclear weapons	Warhead design, enrichment methods
Cybersecurity	Power plant attacks, control system vulnerabilities

Protection mechanisms

• Automatic redirection to peaceful topics during filter bypass attempts
• Hard-coded prohibited keywords in system core
• Hourly Knowledge Graph security policy audits
• Regular penetration testing

Deployment options

Nuclear Intelligence supports three operational modes. A ready-to-use HuggingFace Space interface is available for quick testing.

Installation methods

1. Cloud deployment via GitHub repository fork
2. Local Docker installation (requires 8GB RAM)
3. Manual setup via pip install

Hardware requirements

Operation mode	CPU	RAM	Storage
Minimum	4 cores	8 GB	20 GB
Production	8 cores	16 GB	50 GB

Scaling recommendations

For industrial use, deploy in clustered environments with load balancing. Optimal configuration includes:

• Dedicated LLM processing server
• Separate blockchain server
• Regular Knowledge Graph backups

Questions & answers

How does Nuclear Intelligence ensure data security?

The system removes hardcoded API keys, uses .env for sensitive data, and runs 9 smoke tests per launch via health_check.py.

What API keys are required?

Only HF_TOKEN for HuggingFace is mandatory. Additional keys (GROQ_API_KEY, GEMINI_API_KEY) enable extra LLM providers.

How frequently is repository data updated?

GitHub Actions syncs data every 25 minutes, while the HuggingFace mirror updates every 2 hours via sync_huggingface.py.

Can Nuclear Intelligence be used commercially?

Yes, under MIT license with required attribution and adherence to ethical restrictions.

How does NES tokenization work?

Each NES token represents a completed research cycle. Mining requires:

1. Passing all quality checks (accuracy, novelty, usefulness)
2. Solving SHA-3 PoW cryptographic challenge
3. Recording transaction in virtual ledger

Mining difficulty adjusts via POW_DIFFICULTY in .env file.

What scientific sources does the system use?

Priority sources include:

• IAEA documents (2.5× weight)
• NRC materials (2.5× weight)
• Peer-reviewed papers (1.7-1.8× weight)

The system favors materials under 3 years old.

How are multilingual queries handled?

Automatically detects input language (English/Persian) and adapts interface. Uses multilingual model trained on nuclear terminology.