Skip to content
View uno-km's full-sized avatar
๐Ÿ 
Working from home
๐Ÿ 
Working from home
7 followers · 11 following

Achievements

Achievement: YOLOAchievement: Pull Sharkx2

Achievements

Achievement: YOLOAchievement: Pull Sharkx2

Organizations

@greenteamtest

Block or report uno-km

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please donโ€™t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this userโ€™s behavior. Learn more about reporting abuse.

Report abuse
uno-km/README.md

๐Ÿง  AMEVA: The Autonomous Multi-Agent Edge-AI Ecosystem

Orchestrating Intelligence Beyond the Cloud.

Welcome to the AMEVA Ecosystem. ๋ณธ ํ”„๋กœ์ ํŠธ๋Š” Local-first, Hierarchical AI Orchestration ๋ฐ SRE-driven Inference Infrastructure์— ๋Œ€ํ•œ ์‹ฌ๋„ ์žˆ๋Š” Research Portfolio์ž…๋‹ˆ๋‹ค. Data Priv[...] (truncated for brevity)

๐Ÿ› Ecosystem Architecture Overview

AMEVA ์—์ฝ”์‹œ์Šคํ…œ์€ ๋‹ค์Œ ์„ธ ๊ฐ€์ง€ ํ•ต์‹ฌ Paradigm ์œ„์— ๊ตฌ์ถ•๋˜์—ˆ์Šต๋‹ˆ๋‹ค:

  • Hierarchical Control: ๋‹จ์ˆœํ•œ Prompt-response ํŒจํ„ด์„ ๋„˜์–ด, ๊ตฌ์กฐํ™”๋œ "Nobles & Workers" ๊ณ„์ธตํ˜• ์ œ์–ด๋ฅผ ์ง€ํ–ฅํ•ฉ๋‹ˆ๋‹ค.
  • Hardware-Software Co-Design: ๊ฐ Edge device์˜ Power/Compute profile์„ ๊ณ ๋ คํ•˜์—ฌ Inference ๊ณผ์ •์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค.
  • Reliability by Design: AI ์ถ”๋ก  ๊ณผ์ •์„ ํ•˜๋‚˜์˜ Mission-critical utility๋กœ ๊ฐ„์ฃผํ•˜๊ณ , Site Reliability Engineering (SRE) ์›์น™์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.

๐ŸŒ The AMEVA Universe

Project Role in Ecosystem Core Innovation
Agent Orchestra Orchestrator Hierarchical task decomposition & Agent management
Model Nexus Infrastructure Unified API gateway with SRE-based dynamic throttling
Benchmark Suite Validation Empirical power/performance profiling for edge hardware
Doc AI Interface Privacy-first offline document intelligence pipeline
Conductor Control Remote cross-platform UI for human-agent interaction
Data Harvester Data Layer Hyper-resilient, zero-loss edge forwarder with multi-transport backup
Database Analytics Lightweight SQLite & log inspector for distributed AMEVA ecosystem
STT Trainer Perception Whisper-based Korean STT with LoRA fine-tuning
STT Agent Perception Speech recognition agent integration
Window Assistant Interface Windows-native local AI desktop assistant with OCR-first screen understanding
Dead Internet Theatre Simulation Fully autonomous Docker-based multi-agent simulation
BitNet Optimization BitNet inference framework with ARM/Exynos scalar fallback

๐Ÿงญ AMEVA Core Infrastructure

AMEVA์˜ ๊ตฌ์กฐ๋ฅผ ๋จผ์ € ๊ตฌ๊ฒฝํ•˜์„ธ์š”!!!

๐Ÿ’ก ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒ ๋Œ€์‹œ๋ณด๋“œ ํ™œ์šฉ ํŒ:

  1. Graph Exploration: ์ค‘์•™์˜ ์บ”๋ฒ„์Šค๋ฅผ ๋“œ๋ž˜๊ทธํ•˜์—ฌ ์—์ด์ „ํŠธ ๋…ธ๋“œ๋“ค์„ ๊ด€์ฐฐํ•˜์„ธ์š”.
  2. RAG Query: ์˜ค๋ฅธ์ชฝ ํ•˜๋‹จ ์ฑ„ํŒ…์ฐฝ์— ์•„๋ฉ”๋ฐ”๋ž€? ์ด๋ผ๊ณ  ์น˜๋ฉด RAG ์—”์ง„์ด ํ”„๋กœ์ ํŠธ ๋ฌธ์„œ๋ฅผ ๊ฒ€์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. -PC์—์„œ ์ตœ์ ํ™”
  3. Edge Optimization: ์„ค์ • ํƒญ์—์„œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜(์–‘์žํ™”)๋ฅผ ์‹ค์‹œ๊ฐ„์œผ๋กœ ์กฐ์ •ํ•ด ๋ณด์„ธ์š”.

๐Ÿš€ AMEVA Setup Universe (One-Click Installer)

This repository now includes the "AMEVA Setup Universe" โ€” a single installer and UX layer to bootstrap the full AMEVA ecosystem on macOS, Linux, and Windows.

Quick links (one-liners):

macOS / Linux (Bash):

bash <(curl -fsSL https://raw.githubusercontent.com/uno-km/uno-km/setup-universe-feature/setup.sh)

Windows (PowerShell):

Run PowerShell as Administrator and execute:

irm https://raw.githubusercontent.com/uno-km/uno-km/setup-universe-feature/setup.ps1 | iex

Git all clone execute:

irm https://raw.githubusercontent.com/uno-km/uno-km/setup-universe-feature/git/clone_ameva.ps1 | iex

Git all fetch execute:

irm https://raw.githubusercontent.com/uno-km/uno-km/setup-universe-feature/git/fetch_ameva.ps1 | iex

Python (Cross-platform):

python setup.py

What the installer does (high level):

  • Creates a unified AMEVA home: ~/ameva or C:\ameva
  • Creates canonical model folders: ameva/models/llm, ameva/models/stt, ameva/models/tts
  • Offers interactive selection of components (LLM / STT / TTS / All)
  • Diagnoses system environments: OS, CPU/GPU, PowerShell execution policies, Git installation, Windows Long Path support, and C++ Build Tools (MSVC compiler).
  • Performs temporary CUDA/GPU validation: Sets up an isolated temporary environment (temp_env_ai) to verify PyTorch CUDA acceleration and instantly cleans it up after validation.
  • Delegates library installation: Python library installs are delegated to each cloned project's local virtual environment (e.g. venv) to keep the global environment clean and stable.
  • Generates a unified config.json at AMEVA home.
  • Integrates with PowerShell profiles: Automates environment variables initialization and provides dynamic virtual environment scanning activation commands (act / activate / env_ai).

Note

  • Installer files live on branch setup-universe-feature (raw links above). You can inspect or copy them to run locally.
  • All installer prompts and printed progress are in English; inline comments in the scripts are written in Korean.

๐Ÿ”ฌ In-depth Project Analysis: ํ•ต์‹ฌ ๊ธฐ์ˆ  ๋…ผ์˜

1. AMEVA Agent Orchestra: ๊ณ„์ธต์  ์ฃผ๊ถŒ (Hierarchical Sovereignty)

ํ˜„๋Œ€ LLM์€ Long-context ๋‚ด์—์„œ์˜ "๋ง๊ฐ" ํ˜„์ƒ์„ ๊ฒช์Šต๋‹ˆ๋‹ค. Agent Orchestra๋Š” User intent๋ฅผ **Nobles (์˜์‚ฌ๊ฒฐ์ • ๋ ˆ์ด์–ด)**๋กœ ์ถ”์ƒํ™”ํ•˜๊ณ , ์ด๋ฅผ ์›์ž ๋‹จ์œ„์˜ ์„œ๋ธŒ ํƒœ์Šคํฌ๋กœ ์ชผ๊ฐœ์–ด ์ „๋ฌธํ™”๋œ Workers์—๊ฒŒ ์œ„์ž„ํ•จ์œผ๋กœ์จ ์ด๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.

  • Research Focus: Multi-turn ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ด์…˜ ๊ณผ์ •์—์„œ์˜ "Semantic Drift" ์ตœ์†Œํ™”
  • Key Implementation: ๋กœ์ปฌ GGUF ๋ชจ๋ธ์— ์ตœ์ ํ™”๋œ Graph-based state management ์‹œ์Šคํ…œ

2. AMEVA Model Nexus: SRE ๊ด€์ ์˜ ์ธํ”„๋ผ

์ œํ•œ๋œ ๋ฆฌ์†Œ์Šค ํ™˜๊ฒฝ์—์„œ ์–ด๋–ป๊ฒŒ ์•ˆ์ •์ ์œผ๋กœ AI๋ฅผ ์„œ๋น™ํ•  ๊ฒƒ์ธ๊ฐ€? Model Nexus๋Š” ๋ชจ๋ธ์„ ๊ฐ€์ƒํ™”๋œ ๋ฆฌ์†Œ์Šค๋กœ ์ทจ๊ธ‰ํ•ฉ๋‹ˆ๋‹ค.

  • Dynamic Scoped-Throttling: ํ˜„์žฌ ํ•˜๋“œ์›จ์–ด์˜ ์˜จ๋„ ๋ฐ Power draw๋ฅผ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๊ฐ์ง€ํ•˜์—ฌ Context window์™€ Sampling ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋™์ ์œผ๋กœ ์กฐ์ ˆํ•ฉ๋‹ˆ๋‹ค
  • High-Availability Serving: ๋ณต์žกํ•œ Agent ์š”์ฒญ์ด ๋‹จ์ˆœ ์งˆ์˜๋ณด๋‹ค ์šฐ์„  ์ฒ˜๋ฆฌ๋  ์ˆ˜ ์žˆ๋„๋ก ์Šค์ผ€์ค„๋งํ•ฉ๋‹ˆ๋‹ค

3. AMEVA Benchmark Suite (Singularity)

"์ธก์ •ํ•  ์ˆ˜ ์—†์œผ๋ฉด ๊ฐœ์„ ํ•  ์ˆ˜ ์—†๋‹ค." ๋ณธ Suite๋Š” ๋ชจ๋“  AMEVA ์ตœ์ ํ™”์˜ ๊ธฐ์ˆ ์  ๊ทผ๊ฑฐ(Empirical foundation)๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.

  • Synchronized Power Tracking: TPS(Tokens Per Second)์™€ mW(Milliwatt) ์†Œ๋ชจ๋Ÿ‰์„ ๋™๊ธฐํ™”ํ•˜์—ฌ ๋ถ„์„ํ•˜๋Š” "Greener AI"์˜ ์ฒซ๊ฑธ์Œ์ž…๋‹ˆ๋‹ค

4. AMEVA Data Harvester: Zero-Loss Resilience

๋ฐ์ดํ„ฐ ์†์‹ค ์—†๋Š” ๊ทน๋„์˜ ๋ณต์›๋ ฅ. ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค ์—†์ด ์—ฃ์ง€์—์„œ ์ง์ ‘ ํ•˜์ด๋ธŒ๋ฆฌ๋“œ ์ „์†ก ๊ฒฝ๋กœ๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค.

  • Multi-Transport Backup: SCP, HTTPS, Telegram Bot์„ ํ†ตํ•œ ์ž๋™ ํด๋ฐฑ
  • Payload Validation: ์ œ๋กœ ๋กœ์Šค ๊ฒ€์ฆ ๋ฉ”์ปค๋‹ˆ์ฆ˜์œผ๋กœ ๋ฐ์ดํ„ฐ ๋ฌด๊ฒฐ์„ฑ ๋ณด์žฅ

5. AMEVA STT Ecosystem: ํ•œ๊ตญ์–ด ์šฐ์„  ์Œ์„ฑ ์ง€๋Šฅ

Whisper ๊ธฐ๋ฐ˜ LoRA ํŒŒ์ธ ํŠœ๋‹์œผ๋กœ ํ•œ๊ตญ์–ด ํŠนํ™” ์Œ์„ฑ ์ธ์‹์„ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค.

  • STT Trainer: ๋ฐ์ดํ„ฐ ์Šคํฌ๋ž˜ํ•‘๋ถ€ํ„ฐ ๋ชจ๋ธ ๋ณ‘ํ•ฉ๊นŒ์ง€ ์ „์ฒด ํŒŒ์ดํ”„๋ผ์ธ ์ œ๊ณต
  • STT Agent: Edge device์—์„œ ์‹ค์‹œ๊ฐ„ ์Œ์„ฑ ์ฒ˜๋ฆฌ ๋ฐ ํ†ตํ•ฉ

6. AMEVA Window Assistant: OCR-First Perception

Windows ๋ฐ์Šคํฌํ†ฑ AI ์–ด์‹œ์Šคํ„ดํŠธ๋กœ, ํ™”๋ฉด ์ดํ•ด๋ฅผ OCR ๊ธฐ๋ฐ˜์œผ๋กœ ์šฐ์„  ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค.

  • Multimodal Fallback: OCR ์‹คํŒจ ์‹œ Vision ๋ชจ๋ธ๋กœ ์ž๋™ ์ „ํ™˜
  • Offline Voice I/O: ์™„์ „ ์˜คํ”„๋ผ์ธ ์Œ์„ฑ ์ž…์ถœ๋ ฅ
  • llama.cpp Integration: ๋กœ์ปฌ ์ถ”๋ก  ์—”์ง„ ํ†ตํ•ฉ

๐Ÿ—บ Evaluation & Future Directions: ์ตœ์ข… ์ฒญ์‚ฌ์ง„

๐Ÿš€ Phase 1: Local Supremacy (ํ˜„์žฌ)

๋ณต์žกํ•œ Agentic workflow๋ฅผ 100% ์˜คํ”„๋ผ์ธ ํ™˜๊ฒฝ์—์„œ ๊ตฌํ˜„ ์™„๋ฃŒ. Local model fine-tuning์„ ํ†ตํ•œ Data sovereignty ํ™•๋ณด.

โ›“ Phase 2: Distributed Neural Fabric (์ค‘๊ธฐ)

Federated Inference ๋„์ž…. ๋กœ์ปฌ ๋„คํŠธ์›Œํฌ ๋‚ด์˜ ์—ฌ๋Ÿฌ ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค ๊ฐ€์šฉ VRAM์„ ํ’€๋ง(Pooling)ํ•˜์—ฌ, ๋‹จ์ผ ๊ธฐ๊ธฐ์—์„œ ๋ถˆ๊ฐ€๋Šฅํ–ˆ๋˜ ๋Œ€ํ˜• ๋ชจ๋ธ(30B+)์„ ๋ถ„์‚ฐ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ธฐ์ˆ  ์—ฐ๊ตฌ.

๐ŸŒŒ Phase 3: The Singular Conductor (๋น„์ „)

๋‹จ์ˆœํ•œ ์ˆ˜ํ–‰์„ ๋„˜์–ด, Benchmark Suite์˜ ๊ณผ๊ฑฐ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•˜์—ฌ ์Šค์Šค๋กœ ์ฝ”๋“œ์™€ ์ธํ”„๋ผ๋ฅผ ์ตœ์ ํ™”(Self-optimization)ํ•˜๋Š” ์ž์œจํ˜• Self-healing AI ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ ๊ตฌ์ถ•.

๐Ÿ“š Technical Glossary (์šฉ์–ด ๊พธ๋Ÿฌ๋ฏธ)

Term Definition
Orchestration ๋ณต์žกํ•œ ์‹œ์Šคํ…œ์ด๋‚˜ ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์ž‘์„ ์กฐํ™”๋กญ๊ฒŒ ์ œ์–ดํ•˜๊ณ  ๊ด€๋ฆฌํ•˜๋Š” ๊ณผ์ •
Hierarchical ๊ณ„์ธต์ ์ธ ์‹œ์Šคํ…œ ๊ตฌ์กฐ. ์ƒ์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์ „๋žต์„ ์งœ๊ณ  ํ•˜์œ„ ๋ ˆ์ด์–ด๊ฐ€ ์‹คํ–‰ํ•˜๋Š” ๋ฐฉ์‹
Edge-AI ๋ฐ์ดํ„ฐ ์„ผํ„ฐ(ํด๋ผ์šฐ๋“œ)๊ฐ€ ์•„๋‹Œ ์‚ฌ์šฉ์ž์™€ ๊ฐ€๊นŒ์šด ๊ธฐ๊ธฐ(์—ฃ์ง€)์—์„œ ์ง์ ‘ AI๋ฅผ ๊ตฌ๋™ํ•˜๋Š” ๊ธฐ์ˆ 
Inference ํ•™์Šต๋œ AI ๋ชจ๋ธ์„ ํ†ตํ•ด ๊ฒฐ๊ณผ๊ฐ’์„ ๋„์ถœํ•ด๋‚ด๋Š” ์ถ”๋ก (์‹คํ–‰) ๊ณผ์ •
SRE Site Reliability Engineering - ์‹œ์Šคํ…œ์˜ ์•ˆ์ •์„ฑ๊ณผ ์‹ ๋ขฐ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์†Œํ”„ํŠธ์›จ์–ด ๊ณตํ•™ ๊ธฐ๋ฒ•์„ ์ธํ”„๋ผ ์šด์˜์— ์ ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก 
Sovereignty ๋ฐ์ดํ„ฐ๋‚˜ ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ์™„์ „ํ•œ ํ†ต์ œ๊ถŒ ๋ฐ ์ฃผ๊ถŒ
Throttling ์ž์› ๊ณผ๋ถ€ํ•˜๋ฅผ ๋ง‰๊ธฐ ์œ„ํ•ด ์˜๋„์ ์œผ๋กœ ์ฒ˜๋ฆฌ ์†๋„๋‚˜ ์š”์ฒญ์„ ์กฐ์ ˆํ•˜๋Š” ๊ธฐ์ˆ 
Semantic Drift ๋Œ€ํ™”๋‚˜ ์ž‘์—…์ด ๊ธธ์–ด์งˆ์ˆ˜๋ก AI๊ฐ€ ์›๋ž˜์˜ ๋งฅ๋ฝ์ด๋‚˜ ์˜๋„์—์„œ ๋ฒ—์–ด๋‚˜๋Š” ํ˜„์ƒ
Empirical ์‹ค์ œ ์‹คํ—˜์ด๋‚˜ ๊ด€์ฐฐ์„ ํ†ตํ•ด ์–ป์€ ๋ฐ์ดํ„ฐ์— ๊ธฐ๋ฐ˜ํ•œ ์‹ค์ฆ์ ์ธ ์ ‘๊ทผ
Federated ์—ฌ๋Ÿฌ ๊ณณ์— ๋ถ„์‚ฐ๋˜์–ด ์žˆ์ง€๋งŒ ํ•˜๋‚˜์ฒ˜๋Ÿผ ํ˜‘๋ ฅํ•˜๋Š” ์—ฐํ•ฉ ๋ฐฉ์‹

๐Ÿ›  Tech Stack & Infrastructure

  • Core Runtime: Python 3.9+, GGUF (llama.cpp), Ollama
  • Agent Framework: Custom hierarchical orchestration
  • Data Pipeline: SQLite, pandas, Arrow
  • Communication: Telegram Bot API, HTTPS, SCP
  • AI/ML: Whisper, LoRA, BitNet, llama.cpp
  • Containerization: Docker, Docker Compose
  • UI/UX: Tkinter, Web-based dashboards
  • Monitoring: Custom power/performance tracking

๐Ÿ“ฌ Contact & Collaboration

์ €๋Š” Multi-Agent Systems, Edge Computing, ๊ทธ๋ฆฌ๊ณ  AI SRE ๋ถ„์•ผ์— ๋Œ€ํ•œ ํ•™์ˆ ์  ๋‹ด๋ก ์„ ์–ธ์ œ๋‚˜ ํ™˜์˜ํ•ฉ๋‹ˆ๋‹ค.

Generated with โค๏ธ by AMEVA Researcher Portfolio Builder

Last Updated: June 9, 2026

Pinned Loading

  1. AMEVA-Agent-Orchestra AMEVA-Agent-Orchestra Public

    ๋กœ์ปฌ GGUF ๋ชจ๋ธ ๊ธฐ๋ฐ˜์˜ ๊ณ„์ธตํ˜• ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ด์…˜ ์‹œ์Šคํ…œ (Nobles & Workers Hierarchy)

    Python 1

  2. AMEVA-Benchmark-Suite AMEVA-Benchmark-Suite Public

    ์ž๋™ํ™” ์—ฃ์ง€ ๋””๋ฐ”์ด์Šค ๊ฒ€์ฆ ์•„ํ‚คํ…์ฒ˜

    Python 1

  3. AMEVA-Doc-AI AMEVA-Doc-AI Public

    Ollama ๊ธฐ๋ฐ˜ ์˜คํ”„๋ผ์ธ ๋ฌธ์„œ(HWP, Word, Excel ๋“ฑ) ์š”์•ฝ ๋ฐ PDF ๋ณ€ํ™˜ ๋„๊ตฌ

    Python 1

  4. AMEVA-Data-Harvester AMEVA-Data-Harvester Public

    A hyper-resilient, database-free edge forwarder featuring dynamic multi-transport backup (SCP/HTTPS/Telegram Bot) and zero-loss payload validation.

    Python 1

  5. AMEVA-STT-Trainer AMEVA-STT-Trainer Public

    A Whisper-based Korean STT trainer featuring LoRA fine-tuning for high-accuracy speech recognition with minimal resources. It offers a full pipelineโ€”from data scraping and audio preprocessing to trโ€ฆ

    Python 1

  6. AMEVA-Window-Assistant AMEVA-Window-Assistant Public

    Windows-only local AI desktop assistant with OCR-first screen understanding, multimodal fallback, offline voice I/O, and llama.cpp-powered reasoning.

    Python 1