We built our own LLM where each word turns into a vector via Token Embedding, then processed by Multi-Head Self-Attention to weigh meaning, question, and answer. Output logits go through Softmax with temperature to predict the next word.

1. Token Embedding + Attention + Fine-Tuned Gemini 2

Each token gets a vector. Parallel attention heads process it into logits. We fully replaced Gemini 2’s weights and fine-tuned from scratch — no pre-trained layers used.

2. Vision + Action via ViT + Reinforcement Learning

Images are split into patches, embedded, and passed through attention. The agent self-trains via games (e.g., Minecraft), adapting to new tasks without re-training — thanks to internal self-evolving mechanisms and A2A + MCP architecture.