init: project structure with LAION scorer, conda setup, and research docs

docs/research-commercial-tools.md

@@ -0,0 +1,29 @@
+# Commercial AI Photo Culling Tools — For Reference
+
+These are not self-hosted but represent the state-of-the-art in wedding photo culling.
+
+## FilterPixel DeepCull (2026)
+
+- **Type:** Cloud-based, genre-specific AI
+- **Speed:** 2:58 for 3,000 photos (cloud)
+- **Features:** 10-parameter scoring (emotion, sharpness, composition, storytelling, etc.), memory-based learning, genre-specific models
+- **Accuracy:** 85-95% agreement after training on user's past culling decisions
+- **Price:** $9.99-$19/month
+- **Best for:** Professional photographers needing same-day delivery
+
+## Aftershoot
+
+- **Type:** Local desktop application
+- **Speed:** 8-12 minutes for 1,000 images (local)
+- **Features:** Duplicate detection, blink detection, blur detection, highlight selection, genre selection (portrait/wedding/events)
+- **Learning:** Adapts to user's editing style over time
+- **Price:** $15/month
+- **Best for:** Wedding photographers with multi-day turnaround
+
+## Imagen AI
+
+- **Type:** Cloud-based editing + culling
+- **Accuracy:** 63.4% (noted as lower than FilterPixel/Aftershoot in independent testing)
+- **Features:** Culling studio, AI editing, pay-per-image pricing
+- **Price:** Pay per image (expensive at scale)
+- **Best for:** Photographers already using Imagen for editing