# AMEVA Ecosystem: Comprehensive Technical Architecture & Full-Text AI Reference (llms-full.txt)

This document contains the complete technical specification, APIs, mathematical models, architectural boundaries, and runtimes of the AMEVA ecosystem.

---

## Table of Contents
1. Executive Identity & Engineering Philosophy
2. AMEVA Acronym & 4-Layer Architecture
3. AMEVA Workstation Web (WebGPU Sovereign Local AI)
4. AMEVA-Forge (Browser-Native WebGPU Autograd Engine)
5. Termux-STT (Integrated On-Device Speech Recognition & Diarization)
6. Termux-Train (Bionic ARM64 DAG Autograd & LoRA Training Engine)
7. Termux-Playwright (Non-Root Chromium CDP Automation)
8. Termux-Diffusion (Bionic ARM64 Native Stable Diffusion)
9. AMEVA Sentinel (Privacy-First Security Observability & Risk Scoring)
10. Termux-BitNet (1.58-bit Ternary Quantization Inference Engine)
11. Termux-LlamaCPP (High-Performance GGUF On-Device LLM Backend)
12. Termux-MCP (Model Context Protocol Edge Hub)
13. Termux-Vision (On-Device Computer Vision & Multimodal Embeddings)
14. AMEVA-Runtime (ARM64 Vulkan 1.3 Subgroup Compute & Adreno GPGPU Engine)
15. Termux-TTS (On-Device Text-to-Speech Neural Vocoder)
16. Termux-AIChain (Deterministic DAG Agent Execution & Distributed P2P Mesh)
17. Termux-Cluster (Edge Fleet Coordination & Telemetry Network)
18. Termux-Privacy (Hardware Fingerprint Shield & Telemetry Mask)
19. SRE Dynamic Throttling & Security Isolation
20. AMEVA Labs (Sovereign Systems Research & Community Portal)

---

## 1. Executive Identity & Engineering Philosophy
- **Founder & Systems Architect**: Eunho Kim (김은호 / @uno-km)
- **Education**: Ajou University Graduate School (M.S. in AI candidate), Hanyang Cyber University (B.S. in Applied Software Eng, 3.94/4.5)
- **Philosophy**: Complete data sovereignty, 100% serverless edge architecture, offline privacy by design.

---

## 2. AMEVA Acronym & 4-Layer Architecture
- **A — Agentic**: Autonomous reasoning and planning intelligence.
- **M — Multi-Agent**: Hierarchical delegation (Nobles & Workers).
- **E — Ecosystem**: Unified cross-platform integration (Web, Desktop, Mobile Termux, Edge).
- **V — Virtualized**: Isolated client-side WASM and Bionic sandboxes.
- **A — Automation**: Zero-server visual pipeline execution.

---

## 3. AMEVA Workstation Web
- **URL**: https://ameva-workstation-web-core.vercel.app/
- **Key Modules**:
  - WebGPU Local AI: Qwen2.5 0.5B, 1.5B, 7B running in-browser via WebLLM/WGSL shaders.
  - 3-Sec MapReduce PDF/DOCX Document Intelligence.
  - In-App Video Trimming (FFmpeg WASM).
  - AI Background Removal (WebAssembly Neural Segmentation).
  - Silence Detection & Waveform Trimmer.

---

## 4. AMEVA-Forge
- **PyPI**: `pip install ameva`
- **Docs**: https://uno-km.vercel.app/lib/forge/
- **Features**:
  - Drop-in PyTorch compatibility (`torch.Tensor`, `nn.Linear`, `nn.LayerNorm`, `optim.SGD`, `loss.backward()`).
  - WGSL 16x16 workgroup tiled matmul shaders.
  - GPU reverse-mode automatic differentiation.
  - Deterministic GC finalizers and zero-leak staging buffer pools.

---

## 5. Termux-STT
- **PyPI / npm**: `pip install termux-stt` | `npm install termux-stt`
- **Docs**: https://uno-km.vercel.app/lib/stt/
- **Features**:
  - Unified 3-engine backend: Whisper.cpp (OpenBLAS NEON), Vosk (Kaldi), Sherpa-ONNX.
  - Pure Python 128d X-Vector K-Means Speaker Diarization.
  - Low memory footprint (<350MB RAM on Android Termux).

---

## 6. Termux-Train
- **PyPI**: `pip install termux-train`
- **Docs**: https://uno-km.vercel.app/lib/train/
- **Features**:
  - Pure Python Autograd Core with pluggable NumPy C-Acceleration.
  - SafeTensors zero-copy parser & serialization.
  - Rotary Position Embeddings (RoPE) & Causal Attention.
  - Low-Rank Adaptation (LoRA) on-device training & weight freezing.

---

## 7. Termux-Playwright
- **PyPI / npm**: `pip install termux-playwright` | `npm install termux-playwright`
- **Docs**: https://uno-km.vercel.app/lib/playwright/
- **Features**:
  - Native Bionic Chromium CDP control without PRoot overhead.
  - Persistent disk ledger (.tp_ledger) for process recovery across crashes.
  - Prototype-safe anti-bot stealth deleting `navigator.webdriver`.
  - eMMC flash memory wear protection via RAM-based caching.

---

## 8. Termux-Diffusion
- **PyPI / npm**: `pip install termux-diffusion` | `npm install termux-diffusion`
- **Docs**: https://uno-km.vercel.app/lib/diffusion/
- **Features**:
  - Dual-Engine (Python & Node.js) on-device Stable Diffusion.
  - bfloat16 ARM NEON assembly acceleration for Samsung Galaxy hardware.
  - Zero-root, zero-PRoot native execution.

---

## 9. AMEVA Sentinel
- **npm**: `npm install @ameva/sentinel@alpha @ameva/sentinel-browser@alpha @ameva/sentinel-risk-core@alpha`
- **Docs**: https://uno-km.github.io/ameva-sentinel/
- **GitHub**: https://github.com/uno-km/ameva-sentinel
- **Features**:
  - Privacy-first client security observability.
  - Deterministic mathematical risk scoring (0-100 scale).
  - Cross-browser verified: Chromium, Firefox, WebKit.
  - Zero raw keystroke or location data ingestion.

---

## 10. Termux-BitNet
- **PyPI**: `pip install termux-bitnet`
- **Docs**: https://uno-km.vercel.app/lib/bitnet/
- **Features**:
  - 1.58-bit ternary weight representation (-1, 0, +1).
  - Addition-only GEMM operations reducing memory bandwidth by up to 80%.
  - ARM64 NEON vectorized instruction kernels for edge deployment.

---

## 11. Termux-LlamaCPP
- **PyPI**: `pip install termux-llamacpp`
- **Docs**: https://uno-km.vercel.app/lib/llamacpp/
- **Features**:
  - GGUF multi-quantization support (Q4_K_M, Q5_K_M, Q8_0).
  - High-efficiency context window management with mmap weight paging.
  - Vulkan compute offload for capable mobile GPUs.

---

## 12. Termux-MCP
- **PyPI / npm**: `pip install termux-mcp` | `npm install termux-mcp`
- **Docs**: https://uno-km.vercel.app/lib/mcp/
- **Features**:
  - Standardized Model Context Protocol client/server implementation.
  - Tool registration and dynamic execution sandbox for autonomous agents.
  - JSON-RPC over stdio and HTTP/SSE transport modes.

---

## 13. Termux-Vision
- **PyPI**: `pip install termux-vision`
- **Docs**: https://uno-km.vercel.app/lib/vision/
- **Features**:
  - Lightweight MobileNet and Vision Transformer feature extractors.
  - Fast cosine similarity indexing for on-device visual search.
  - Zero cloud dependency for optical classification.

---

## 14. AMEVA-Runtime
- **PyPI**: `pip install ameva-runtime`
- **Docs**: https://uno-km.vercel.app/lib/vulkan/
- **Features**:
  - Direct SPIR-V compute pipeline generation.
  - Vulkan 1.3 subgroup ballot and arithmetic operations.
  - Qualcomm Adreno 650/830 GPU defect workarounds and LDS bank conflict minimization.

---

## 15. Termux-TTS
- **PyPI / npm**: `pip install termux-tts` | `npm install termux-tts`
- **Docs**: https://uno-km.vercel.app/lib/tts/
- **Features**:
  - Neural text-to-speech vocoder running in low-resource environments.
  - Streaming audio generation with PCM chunk buffering.
  - Support for multi-lingual synthetic output.

---

## 16. Termux-AIChain
- **PyPI**: `pip install termux-aichain`
- **Docs**: https://uno-km.vercel.app/lib/aichain/
- **Features**:
  - Deterministic execution graphs with checkpointing and state replay.
  - P2P message passing across edge swarm nodes.
  - Protocol-level context compression.

---

## 17. Termux-Cluster
- **PyPI**: `pip install termux-cluster`
- **Docs**: https://uno-km.vercel.app/lib/cluster/
- **Features**:
  - Work-stealing distributed task scheduling for edge devices.
  - Fleet health monitoring and heartbeat signaling.
  - Automatic node discovery and failover routing.

---

## 18. Termux-Privacy
- **Docs**: https://uno-km.vercel.app/lib/privacy/
- **Features**:
  - Client fingerprint masking and synthetic hardware profile generation.
  - Zero-knowledge data sanitization before external egress.
  - Local entropy injection to resist canvas/WebGL tracking.

---

## 19. SRE Dynamic Throttling & Security Isolation
- **Security Guarantee**: AMEVA tools execute within safe client-side WASM or unprivileged user-space Termux sandboxes with zero root elevation.
- **SRE Resilience**: Dynamic throttling based on device temperature, memory headroom, and battery state.

---

## 20. AMEVA Labs (Sovereign Systems Research & Community Portal)
- **Portal**: https://uno-km.vercel.app/labs/
- **Architecture**:
  - High-density academic disclosure feed with TechArticle Schema.org metadata and Citation JSON-LD for AI crawler retrieval (GPTBot, ClaudeBot, PerplexityBot).
  - Tree-structured hierarchical comment board with IPv4 subnet masking (`123.45.***.***`) and cryptographic salt hashing.
  - In-browser Workstation lightweight Markdown/TeX drafting studio with live client-side preview.
  - Multi-tiered board categories: `news` (official dispatches), `anal` (in-depth silicon/kernel benchmarks), `blog` (curated technical essays), `board` (sovereign open discussion).
  - Client-side telemetry radar: WebGL GPU renderer fingerprinting, 2D Canvas hash, active viewport gaze tracking, and millisecond dwell time metrics.