Leavesfly/TinyAI

What it solves

TinyAI is a full-stack AI framework built entirely in Java, designed to eliminate the dependency on Python and complex deep learning environments. It provides a comprehensive technical stack that bridges the gap between low-level numerical computation and high-level AI applications, making it suitable for both educational purposes and enterprise-grade Java integration.

How it works

TinyAI uses a six-layer architecture to organize its 23 core modules:

1. Base Layer: A custom multi-dimensional array library for tensor operations.
2. Engine Layer: An automatic differentiation engine with dynamic computation graphs.
3. Framework Layer: Core machine learning tools, neural network layers (CNN, RNN, Transformer), and reinforcement learning algorithms.
4. Embodied Intelligence Layer: Tools for autonomous driving simulation, robot control, and world models.
5. Model Layer: Implementations of LLMs including the GPT series, DeepSeek R1/V3, and multimodal models like Banana.
6. Agent Layer: A framework for RAG, multi-agent collaboration, and cognitive patterns (ReAct, CoT).

Who it’s for

• Java Developers: Those who want to integrate AI into Java ecosystems without switching to Python.
• AI Students: Learners seeking a transparent, well-documented (300k+ words) framework to understand AI from the ground up.
• Enterprise Architects: Developers needing a modular, scalable AI system with production-grade reliability and DevOps support.

Highlights

• Pure Java Implementation: Zero Python dependencies, requiring only JDK 17+.
• Broad Scope: Covers everything from basic tensors to advanced DeepSeek R1/V3 reasoning models and VLA (Vision-Language-Action) architectures.
• Educational Depth: Includes over 890 test cases and a structured learning path with extensive documentation.
• Advanced AI Features: Supports Mixture of Experts (MoE), RAG, and "imagination training" for embodied intelligence to improve sample efficiency.