One script becomes three click-worthy thumbnails in 30–60 seconds
Working with marketing agency Company R, we built an automatic YouTube-thumbnail generator and jointly delivered a proposal for an integrated customer-management system for Company Y, a golf-specialist travel agency.
- Client
- Company R (marketing agency)
- Period
- Jan–Feb 2026
- Role
- AI production-tool development + joint proposal
30–60 seconds
Generation time
3–5
Outputs per run
1280×720
Output resolution
The Problem
- The thumbnail-production bottleneck, where every video had to pass through a designer's hands
Approach
- A dual-AI pipeline — text analysis and copy (Claude) plus image generation (Gemini) — driven by a 6-layer prompt engine
- On photo upload, background removal and face-orientation detection auto-place the subject
- Built-in automatic validation rules (Zero-Rules) prevent text corruption and duplicated subjects
Systems Built
| Module | Description |
|---|---|
| 3-step wizard UI | Input → preview → generate |
| 3–5 thumbnail variants per run | ZIP download |
Tech Stack
FastAPIStreamlitClaudeGeminiPillowrembgOpenCV
