KJP AI Integration — Feature Brief

Consolidated from Notion 2026-04-23. Internal use, can be adapted for client-facing. Status: Phase 2 — deferred to fall/winter 2026. Preserved for when work reactivates.

What the client asked for, what we're building, what's still open, and a plain-English overview of the process.

1. What the client asked for

• Auto-generate tags, color, and descriptions for ~8,400 photos that are missing them
• Ongoing — runs on new uploads too, not just a one-time backfill
• Always review before anything goes live

2. What the finished system does

• Analyzes each photo with AI vision
• Tags: AI picks from predefined tag list (classification, not generation) — existing taxonomy is the candidate pool. Adds mood tags (uplifting, calm, soothing) and species/subject tags where missing.
• Color: AI selects from the existing 16-term color taxonomy — same classification approach, supports multi-color.
• Descriptions (bigger lift, separate decision): Free-text generation — needs voice/tone matching, more review, higher cost. Only pursue if SEO value confirmed.
• Queues all output for KJP review → approved items push to WP taxonomies.
• Must be custom-built (images aren't in WP Media Library — off-shelf plugins won't work).
• Pre-req: taxonomy cleanup pass first (typos, duplicates in tags + color).

How It Works — Process Overview

1. Clean up the existing taxonomy. Before the AI runs, prepare tag and color data so it has clean options to work from. - Remove duplicate and misspelled tags (e.g., "Pruple," "Yelllow") - Clear out unused categories that have never been assigned to a photo - Confirm the final list of color and mood tags the AI will use

2. AI analyzes each photo. A custom tool pulls each photo from the website and sends it to an AI vision model, which reviews the image and makes selections. - Picks the most relevant tags from the existing tag list - Assigns a color (or colors) from the 16-color category list - Rates its own confidence so lower-confidence results get flagged for closer review

This is the step where AI processing fees apply.

3. Start with a small test batch. Before processing the full catalog, run a pilot of ~50 photos. - Client reviews AI's suggestions and tells us what's right and what's off - Fine-tune based on feedback - Repeat until satisfied with quality before scaling up

4. Client reviews and approves before anything goes live. All AI suggestions land in a simple review dashboard — nothing touches the website until approved. - See the photo alongside its suggested tags and color - Approve, edit, or skip each one - Changes are only pushed to the site after sign-off

5. Approved updates publish to the website. Once batch is approved, everything syncs automatically. - Tags and colors appear on the correct photos - No manual data entry required on client end

6. New photos are handled automatically going forward. After the initial catalog is complete, the system continues running on its own. - New imports are picked up automatically - They're analyzed and added to the review queue within 24 hours

3. Open questions — decisions before build

• Descriptions: in scope or just tags + color? (tags/color = easier; descriptions = bigger lift)
• Mood tag taxonomy: need to define the predefined list before AI can pick from it
• Multi-color: tag all prominent colors or just dominant?
• Budget/timeline expectations