CAID: Computer AI Design

Breaking the Bottleneck of Hardware Innovation through Agentic AI

---

"CAID는 하드웨어 설계의 패러다임을 바꾸기 위한 오픈 소스 프로젝트입니다. 기존의 CAD가 인간의 설계를 '보조'하는 도구였다면, CAID는 LLM과 World Model을 결합하여 AI가 직접 설계의 주체(Primary Designer)가 되는 세상을 꿈꿉니다."

---

Why CAID?

오늘날 소프트웨어의 발전 속도에 비해 하드웨어 혁신은 느립니다. 그 중심에는 '수동 모델링'이라는 병목 현상이 있습니다. 숙련된 엔지니어가 수백 시간을 들여 직접 도면을 그려야 하는 제약을 극복하기 위해, CAID는 자연어(Text)만으로 3D 부품과 복잡한 어셈블리를 생성하는 것을 목표로 합니다.

Key Features

• Generative Parametric Design: 단순한 형상이 아닌, 수정 가능한 파라메트릭 코드(CadQuery)를 생성합니다.
• Agentic Multi-Agent System:
  • Architect (agents/architect.py): 설계 의도 분석 및 전체 전략 수립
  • Designer (agents/designer.py): 세부 부품 모델링 코드 생성
  • Critic (agents/critic.py): 제조 가능성(DFM) 및 물리적 타당성 검토
• Physics-Aware Modeling: FEA 결과(sim/fea_engine.py)와 공차 해석(sim/tolerance.py)을 설계에 자동 반영합니다.
• REST API: FastAPI 기반 엔드포인트(api/routes.py)로 외부 툴 연동이 가능합니다.

Repository Structure

agents/          AI 에이전트 (Architect, Designer, Critic)
api/             FastAPI REST 엔드포인트
core/            LLM 래퍼, 오케스트레이터, 세션 관리, 스키마
geometry/        CadQuery / OpenSCAD 커널 연동
prompts/         Jinja2 프롬프트 템플릿
sim/             FEA, 공차(Tolerance), 시뮬레이션 서비스
examples/        실전 적용 케이스
output/          생성된 모델 파일 출력 디렉터리

---

CAID Roadmap: The Modular Synthesis Path

✅ Phase 1 & 2: Core & Assembly (Completed)

• Natural language → parametric 3D geometry (CadQuery)
• Agentic self-correction loop with DFM critique (up to 3 iterations)
• Lightweight FEA and tolerance stack-up analysis (sim/)
• 2-part assembly with interference detection
• FastAPI REST interface (api/routes.py)
• CLI entrypoint (caid_cli.py)

---

🚀 Phase 3: Smart Part Library (Current Focus)

Goal: Turn every generated model into a reusable "Brick" that future agents can retrieve, compose, and adapt — eliminating redundant generation and enabling cumulative design knowledge.

Task	Module	Description
Part Repository	library/repository.py	SQLite-backed store: save/load/query parts by name, tags, parameters
Functional Metadata	library/metadata.py	AI-extracted annotations: mounting holes, mating axes, bounding box, material
Semantic Search	library/search.py	Embedding-based similarity search over part descriptions
Structural Sanity Check	sim/quick_check.py	Rule-based pre-FEA viability filter (wall thickness, overhang angle)
Library API	api/routes.py	GET /parts, GET /parts/{id}, POST /parts REST endpoints

First Task → See "Getting Started: Phase 3" below.

---

🛠️ Phase 4: Advanced Modular Assembly (Planned)

Goal: Assemble multi-part systems from natural language commands, with automatic constraint resolution.

• Semantic Mating: Text commands like "Attach A to B via M3 holes" → constraint solver maps named features
• Global Tolerance Control: Automatic clearance/interference adjustment across the full assembly tree
• Standard Part Integration: Off-the-shelf catalogue (bolts, bearings, motors) pulled from the part library
• Assembly Graph: DAG representation of part relationships enabling recursive sub-assembly reuse

---

🌌 Phase 5: Autonomous System Building (Vision)

Goal: Given a high-level functional spec, CAID decomposes, designs, simulates, and validates a complete mechanical system with minimal human intervention.

• Recursive Synthesis: Complex machines built from modular sub-assemblies retrieved from the Phase 3 library
• Generative Layout: AI proposes optimal part arrangements to satisfy spatial, load, and thermal constraints
• World Model Integration (core/world_model.py): Physics priors guide the generative loop end-to-end
• Closed-Loop Validation: Simulation results automatically trigger re-design until specs are met

---

Getting Started: Phase 3 — Part Library

The first concrete task for Phase 3 is implementing library/repository.py: a lightweight, queryable store for generated parts.

What to build

# library/repository.py
@dataclass
class PartRecord:
    id: str                  # UUID
    name: str
    description: str
    tags: list[str]
    parameters: dict         # e.g. {"width": 40, "height": 20, "material": "PLA"}
    cadquery_code: str        # the generated script
    stl_path: str | None     # path under output/
    created_at: str          # ISO-8601

class PartRepository:
    def save(self, part: PartRecord) -> str: ...
    def get(self, part_id: str) -> PartRecord | None: ...
    def search(self, query: str, tags: list[str] = []) -> list[PartRecord]: ...
    def list_all(self) -> list[PartRecord]: ...

Integration points

1. core/orchestrator.py — call repo.save() after every successful generation
2. api/routes.py — expose GET /parts and POST /parts using PartRepository
3. agents/architect.py — call repo.search() before generating; reuse or adapt existing parts

Suggested storage

• SQLite via sqlite3 (stdlib, no extra dependency): one table parts, JSON columns for tags and parameters
• STL files remain under output/; the DB stores the path reference

---

How to Contribute

CAID는 인류의 제작 능력을 확장하고자 하는 모든 '호모 파베르(Homo Faber)'를 환영합니다. 이슈 제안, 코드 기여, 아이디어 공유 등 무엇이든 좋습니다.

License

This project is licensed under the MIT License.