01 / AI WORKFLOW TOOL
BLD8 Workbench
A planned tool for inspecting the model calls, sources, tools and cost behind an AI answer.
01 / Current state
Architecture and draft decisions.
The repository contains a proposed architecture and three draft decision records. Application development has not started.
The first use case is answering questions about a public design system. Source material is still being selected.
02 / First version
Start with one model and a request history.
The proposed first version uses Next.js and Node, with sign-in, request history, usage limits and a budget cap.
PLANNED ARCHITECTURE
- InterfaceNext.js / React
- Request handlingNode / AI SDK
- Model + storageOne model / Postgres
Python-based ingestion and retrieval are planned for a later stage, followed by checks against a small reference dataset.
Architecture notes03 / Questions to try
Questions to test.
Draft questions for the initial design-system use case.
- Which component fits a particular interaction, and why?
- Does a component support multi-select?
- What accessibility guidance applies to dialogs?
- Which token controls a particular surface’s spacing?
- Which components are deprecated, and what replaces them?
04 / Draft decisions
Technical decisions.
These records are drafts and may change during implementation.
001 — Hosting and deployment
Where the application and the later retrieval service will run. Read draft ADR 001.
002 — The Node / Python boundary
Which work stays in Node and what belongs in the Python service. Read draft ADR 002.
003 — Tool permissions
Start with read-only tools. An approval flow is needed before tools can change anything outside the app. Read draft ADR 003.