CAD Train

Parametric 3D CAD modelling pipeline for downhole tool components, combining:

• ManifoldCAD — parametric solid modelling in the browser
• Svelte 5 + Threlte — interactive 3D viewers
• Claude vision API — component identification from reference images
• Retrieval-Augmented Generation — persistent training cache that improves with use

Routes

Path	Purpose
/	Landing page with feature cards
/components	Parametric component library (18 primitives) with live 3D + SVG + PNG export
/reverse	Upload an image → VLM identifies component + estimates params → live 3D render
/training	Tabbed viewer for completion tool training data
/tests	Visual test recordings + cache statistics
/tools/bottom-sub	Dedicated Bottom Sub (HAL10408) parametric viewer
/tools/ratch-latch	Dedicated Ratch-Latch Receiving Head viewer
/api/identify	POST: RAG-based image → component identification (Claude + training cache)
/api/refine	POST: Iterative parameter refinement via SSIM + Claude
/api/accept	POST: Append user-validated result to persistent training cache
/api/cache/stats	GET: Training cache statistics

Local development

bun install
bun run dev

Open http://localhost:3333.

Set ANTHROPIC_API_KEY in a .env file for the identification endpoints.

To re-seed the training cache from the primitive training data:

bun run seed

Build and run in Docker

# Build
docker build -t cadtrain .

# Run (cache is in-memory only)
docker run -p 3333:3333 \
  -e ANTHROPIC_API_KEY=sk-ant-... \
  cadtrain

# Run with persistent cache volume
docker run -p 3333:3333 \
  -e ANTHROPIC_API_KEY=sk-ant-... \
  -v $(pwd)/data:/data \
  cadtrain

The container listens on $PORT (default 3333). The persistent cache is stored at /data/cache.jsonl, and is seeded from the baked-in cache on first run.

Deploy to Railway

1. Push this repo to GitHub.
2. In Railway, click New project → Deploy from GitHub and select this repo.
3. Railway detects Dockerfile automatically. The railway.json file sets the health check.
4. Open the service → Variables → add ANTHROPIC_API_KEY.
5. Open the service → Volumes → add a volume mounted at /data (1 GB is plenty).
6. Railway builds and deploys. The public URL will be shown in the service overview.

Health check path: /api/cache/stats — returns { total, bySource, totalUses }.

Architecture overview

┌───────────────┐   ┌───────────────────┐   ┌──────────────────┐
│  Upload image │──▶│  pHash retrieval  │──▶│  Few-shot Claude │
└───────────────┘   │  from cache.jsonl │   │  (top-K examples)│
                    └───────────────────┘   └────────┬─────────┘
                                                     │
                                                     ▼
                                           ┌────────────────────┐
                                           │  Component + params│
                                           └────────┬───────────┘
                                                    │
                                                    ▼
                                           ┌────────────────────┐
                                           │  Live 3D render    │
                                           │  (ManifoldCAD)     │
                                           └────────┬───────────┘
                                                    │
                                                    ▼
                                           ┌────────────────────┐
                                           │  Auto-refine loop  │
                                           │  SSIM → Claude     │
                                           └────────┬───────────┘
                                                    │
                                                    ▼
                                           ┌────────────────────┐
                                           │  Save to cache     │
                                           │  (persistent)      │
                                           └────────────────────┘

• Training cache — JSONL file at training_data/cache.jsonl. Seeded from primitive training data (122 records) and grows with user-accepted identifications.
• Retrieval — perceptual hash (sharp + manual DCT) + Hamming distance. Top-5 neighbors included as few-shot context.
• Image comparison — pure-TS implementation (src/lib/training/image_diff.ts) computing SSIM, pixel diff, and Sobel edge diff. No Python dependency.

Project layout

src/
├── routes/           # SvelteKit routes (pages + API)
├── lib/
│   ├── components/   # Component library + ManifoldCAD builder + SVG exporter
│   ├── tools/        # Bottom Sub + Ratch-Latch dedicated tools
│   ├── training/     # Persistent cache, pHash, image diff
│   ├── shared/       # Shared Svelte components
│   └── viewer/       # Generic builder for batch/training viewer
scripts/
└── seed_cache.ts     # Re-seed cache.jsonl from prim_* training data
training_data/
├── cache.jsonl       # Persistent training index (committed)
└── prim_*/           # Source training data per primitive component