tsaol Image Moderation Pipeline
Cost-optimized image content moderation using Claude's vision capabilities.
Core Strategies
- Image preprocessing — Resize/compress to minimize token usage
- Tiered cascade — pHash → lightweight ML → Haiku → Sonnet
- Prompt caching — Cache moderation rules for 90% token savings
- Batch API — 50% cost reduction for non-realtime workloads
- Structured output — Minimal JSON responses to cut output tokens
Usage
pip install -e .
from image_moderation import ImageModerationPipeline
pipeline = ImageModerationPipeline(
anthropic_api_key="sk-...",
enable_cache=True,
cascade_levels=["prefilter", "haiku", "sonnet"],
)
result = pipeline.moderate("path/to/image.jpg")
print(result)
# ModerationResult(safe=False, category="violence", confidence=0.95, cost=0.0003)
Cost Comparison
| Strategy | Tokens/Image | Monthly Cost (1M images) |
|---|---|---|
| Raw → Sonnet | ~5000 | $$$$$ |
| Resized → Haiku | ~1000 | $$ |
| Full pipeline | ~100 (avg) | $ |