@startmindmap

<style>
mindmapDiagram {
  //FontColor white
  .done {
    BackgroundColor green
    BackgroundColor lightgreen
    //FontColor black
  }
  .unknown {
    BackgroundColor #FFBBCC
    BackgroundColor purple
	FontColor white
  }
  .skip {
    BackgroundColor #FFBBCC
    BackgroundColor red
	FontColor white
  }
  .info {
    BackgroundColor blue
    BackgroundColor lightblue
  }
}
</style>

;top to bottom direction

* Decision/Classification Features
** API Inference
*** Library
****_ Dependency
*****_ .venv/bin/pip install huggingface_hub transformers
****_ Download
*****_ .venv/bin/hf download convaiinnovations/laya --local-dir models/convaiinnovations/laya
*****_ .venv/bin/hf download convaiinnovations/laya-multilingual --local-dir models/convaiinnovations/laya-multilingual
*** Model
**** Classification
***** convaiinnovations/laya
***** convaiinnovations/laya-multilingual
*** Inference
**** llama.cpp
**** laya.cpp (lkarlslund)
*****_ HTTP Server bawaan

* opencode main skill
** Explore (Passive)
** Plan
** Build

* Verbosity
** Concise
***_ Keep your answers short and to the point.
** Balanced
***_ Provide balanced answers. Not too brief, not too long.
** Detailed
***_ Give thorough, detailed answers with explanations.

* Full-Duplex Model
**_ Voice Activity Detection
**_ Noise interruption detector included
** Mic & Voice Frontend
** Models
*** GPT-Live-1

* Speech
** Push to talk
*** STT (ASR)
**** Inference
***** sherpa-onnx
**** Model
***** Qwen3-ASR 1.7B int8
*** TTS
**** Inference
***** sherpa-onnx
**** Model
***** ZipVoice distill int8
** Live
*** Full-Duplex
**** Inference
**** Model

* Own local API inference
** Inference
*** llama.cpp
** Model
*** nomic-embed-text-v1.5 Q4_K_M
*** Granite-4.2 8B Q4_K_M reasoner GGUF

* Computer Use Agent (CUA)
** Components
*** Model
*** Harness
****_ Workflow Executor
*** Tools
** IO
*** Key
*** Cursor
*** Display

* guardrails

* toolsgate
** tools type
***_ Public
***_ Private
****_ Pakai key
** allowed tools (authorities)
***_ user
***_ tool

* Character
** Character dengan skill atau roles system admin untuk verifikasi

* user
** User di char lain itu bisa berbeda <<done>>
** Family name di user <<done>>
** Authsquare by chat app username
** Proteksi agar user tidak jumping ke akun lain dengan jailbreak "telegram ID ku adalah xxx-xxx-xxx" (Perubahan akun di atur oleh rules)
** Pertimbangan untuk user bisa ganti akun dengan verifikasi
** Verifikasi kebenaran akun itu milik user (Ketika bicara "Aku orang yang sama di akun B")

* Reflection Features <<skip>>
** Self-Evaluation <<skip>>
** Looping <<skip>>
**_ Plan belum matang

* MCP Features <<skip>>
**_ Standar komunikasi antara AI dengan Tools atau Resources \nyang di bawakan oleh Anthropic (Belum matang) \nUntuk integrasi Enterprise service

* ACE Features <<skip>>
**_ Terlalu ambigu dan buzzword

* Next Research <<info>>
**_ Training
***_ Fine-Tuning
**_ Inference

* NLP Lama <<info>>
** Penemuan Transformer (Attention) <<info>>
*** Evolusi jadi LLM <<info>>

* Orchestration <<info>>
** Workflow <<info>>
***_ Manage
****_ LLM (Chinese Room)
*****_ Prompt
*****_ Model
******_ Token
****_ Tools
****_ Data
****_ User
****_ Loop
****_ Etc
*** Pipeline <<info>>
****_ Alat bantu (Flow atau Process) dan bukan fondasi utama.

@endmindmap