Hi NVIDIA Developer Community,
I’m sharing an open-source local LLM workflow project I’ve been building:
GitHub:
SCBKR is a local responsibility-chain workbench for LLMs. It is not a model itself. It is designed to sit in front of local or API-based models and control when generation, review, storage, replay, and retrieval are allowed.
The core idea is simple:
Before an LLM generates, the user must confirm a structured responsibility frame.
After generation, the output must be reviewed before it can be stored.
Failed outputs cannot automatically become memory.
Stored successful cases can later be retrieved as advisory references, but retrieval does not auto-confirm, auto-generate, or auto-commit anything.
Current status:
- P13 core is complete.
- SQLite persistence is implemented.
- JSONL replay ledger is implemented.
- Physical storage for approved corpus / logic / exports / signed memory rules is implemented.
- Similar-case retrieval is implemented.
- ChromaDB is optional.
- SQLite deterministic fallback retrieval remains the durable baseline.
- ChromaDB and SQLite fallback candidates are merged and re-scored deterministically.
- P14 desktop packaging is still pending.
Supported direction:
- Local LLM workflows
- Ollama / LM Studio style local model setups
- OpenAI-compatible API connections
- Human-in-the-loop confirmation
- Replayable responsibility chains
- Local memory protection
- Retrieval without automatic decision takeover
My goal is to turn this into a downloadable desktop app where users can connect their own local model or API key, then use SCBKR as a responsibility-chain layer in front of the model.
Planned P14 direction:
- Desktop app packaging
- Sandbox mode without requiring a real model
- One-click local workflow testing
- Optional connection to local LLMs or API endpoints
I’m especially interested in feedback from developers working with local LLMs, RTX AI PCs, NIM workflows, or human-in-the-loop AI safety / reliability systems.
This project is independent and open-source. It is not an official NVIDIA project or endorsement. I’m sharing it here because the local AI PC / NIM / local LLM workflow direction seems aligned with the kind of developer environment this project is built for.
中文說明：
大家好，我想分享一個我正在開發的開源本地 LLM 工作流專案：
GitHub:
SCBKR 是一套本地責任鏈工作台。它本身不是模型，而是放在模型前方，控制模型什麼時候可以生成、什麼內容可以驗收、什麼結果可以入庫、什麼案例可以被回放與檢索。
核心概念很簡單：
模型生成前，使用者必須先確認責任框架。
模型生成後，輸出必須經過驗收。
驗收失敗的內容不能自動變成記憶。
成功入庫的案例之後可以被檢索，但檢索結果只作為參考，不會自動確認、不會自動生成、不會自動入庫。
目前狀態：
- P13 核心已完成。
- 已支援 SQLite 任務保存。
- 已支援 JSONL 回放帳本。
- 已支援通過驗收後的 corpus / logic / exports / signed memory rules 入庫。
- 已支援相似案例檢索。
- ChromaDB 是可選項目。
- SQLite deterministic fallback 是保底檢索來源。
- ChromaDB 與 SQLite fallback 會合併後重新計分排序。
- P14 桌面 App 封裝尚未開始。
專案方向：
- 本地 LLM 工作流
- Ollama / LM Studio 類本地模型接入
- OpenAI-compatible API 接入
- Human-in-the-loop 確認
- 可回放責任鏈
- 本地記憶保護
- 檢索不自動奪取決策權
我的目標是把它做成可下載的桌面 App，讓使用者可以接自己的本地模型或 API key，並用 SCBKR 作為模型前方的責任鏈控制層。
P14 預計方向：
- 桌面 App 封裝
- 不需要真模型也能測試的 Sandbox Mode
- 一鍵本地流程測試
- 可選接本地 LLM 或 API endpoint
我想聽聽正在做 local LLM、RTX AI PC、NIM workflow、human-in-the-loop AI reliability / safety 方向的開發者意見。
這是獨立開源專案，不是 NVIDIA 官方專案或官方背書。我會發在這裡，是因為它與 local AI PC / NIM / local LLM workflow 的方向有關。