@zombay/baynet

Open-source HITL content moderation + AI training data pipeline.

BayNet bridges automated NSFW detection and human oversight. It persists detection results, builds a human review queue, and exports corrections as labeled training data (COCO format) for model improvement.

Install

npm install @zombay/baynet

Quick Start

import { BayNet, nudenetBackend, simpleAuth } from "@zombay/baynet";

const baynet = new BayNet({
  database: yourDatabaseAdapter,  // @zombay/baynet-d1, @zombay/baynet-postgres, or custom
  storage: yourStorageAdapter,    // @zombay/baynet-r2, @zombay/baynet-s3, or custom
  auth: simpleAuth({ isReviewer: (id) => ADMIN_IDS.includes(id) }),
  backends: [nudenetBackend({ url: "https://your-nudenet-service.com" })],
  hooks: {
    onCsamIndicator: async (detection) => {
      // Required if csam_indicators backend is registered
      await reportToAuthorities(detection);
    },
  },
});

// Detect and persist
const { detectionId, results } = await baynet.detectAndPersist({
  imageBase64: "...",
  mimeType: "image/png",
  sourceType: "preview",
});

// Review queue
const queue = await baynet.getReviewQueue({ status: "pending" });

// Submit review with annotations
await baynet.submitReview({
  detectionId: "det_123",
  reviewerId: "admin_1",
  status: "corrected",
  annotations: [
    { type: "false_positive", regionIndex: 0 },
    { type: "missed_detection", correctedClass: "BUTTOCKS_EXPOSED", correctedBox: [100, 200, 50, 60] },
  ],
});

// Export COCO training data
const { manifest, exportId } = await baynet.exportCoco({ exporterId: "admin_1" });

Adapter Interfaces

BayNet uses pluggable adapters. Implement DatabaseAdapter, StorageAdapter, and AuthAdapter for your platform, or use a pre-built adapter:

| Package | Platform | |---------|----------| | @zombay/baynet-d1 | Cloudflare D1 (Drizzle ORM) | | @zombay/baynet-r2 | Cloudflare R2 | | @zombay/baynet-postgres | PostgreSQL (coming soon) | | @zombay/baynet-s3 | AWS S3 (coming soon) |

Detection Backends

• NudeNet — nudenetBackend({ url }) — connects to a NudeNet Flask microservice
• Gemini Safety — geminiSafetyBackend({ apiKey }) — Google Gemini safety classifier (fail-closed)

Features

• 7 detection categories, 33 class labels
• Active learning priority scoring (uncertain predictions reviewed first)
• CSAM hook enforcement (SDK refuses to init without mandatory reporting hook)
• COCO JSON export with human corrections applied
• Batch review (up to 100 detections)
• Audit logging

License

MIT