Automations:
- 2-column layout: WhatsApp phone LEFT, education RIGHT
- Right column: 'How it works' (5 numbered steps), performance stats, timing controls, reply commands, tips
- Hero spans full width with photo+dark panel
- Improvement CTA is a prominent card, not floating text
- No misalignment — phone fills left column naturally
Collect:
- Appeals shown as visible gap-px grid cards (not hidden dropdown)
- Each card shows name, platform, amount raised, pledge count, collection rate
- Active appeal has border-l-2 blue indicator
- Platform integration clarity: shows 'Donors redirected to JustGiving' etc
- Educational section: 'Where to share your link' + 'How payment works'
- Explains bank transfer vs JustGiving vs card payment inline
AI model: Stripped all model name comments from code (no user-facing references existed)
All AI features now use Gemini 2.0 Flash via the existing API key.
Falls back to OpenAI if OPENAI_API_KEY is set instead.
Falls back to heuristics if neither key exists.
Gemini free tier: 15 RPM, 1M tokens/day, 1500 RPD
At PNPL's scale this is effectively unlimited and costs £0.
Changed:
- src/lib/ai.ts: chat() → tries Gemini first, OpenAI fallback
- src/app/api/automations/ai/route.ts: same dual-provider pattern
- docker-compose.yml: GEMINI_API_KEY added to app environment
All 11 AI features now work:
- Smart amount suggestions, message generation, fuzzy matching
- Column mapping, event parsing, impact stories, daily digest
- Nudge composer, donor classification, anomaly detection
- A/B variant generation, rewrites, auto-winner evaluation
