Inspiration
Most AI image tools still feel like one-shot prompt boxes — you describe an image, get a result, and start over for every edit. We wanted a workflow where creators, marketers, and e-commerce teams could chat naturally to generate, refine, and ship production-ready visuals without design software or prompt engineering.
What it does
DeepKolor is a conversational AI design agent at deepkolor.ai. Users describe what they need in plain language and iterate in a continuous chat to:
- Generate product photos, posters, and marketing visuals
- Edit photos (remove people, change expressions, virtual try-on)
- Run 15+ specialized MiniApps (background removal, upscaling, storyboards, comics, and more)
- Export watermark-free images for commercial use
How we built it
DeepKolor is built on TanStack Start (React 19 + Vite + Nitro), with better-auth for authentication, Drizzle ORM for multi-database support, and Cloudflare R2 for asset storage. The AI layer uses the Vercel AI SDK with multiple model providers (Gemini, Replicate, Fal, OpenRouter) and a skill-routing agent that picks the right tool for each request.
The product surface combines:
- A chat-first Media Agent with context-aware editing
- Config-driven MiniApps for focused workflows
- Full SaaS infrastructure: credits, subscriptions, RBAC admin panel, and i18n (EN/ZH)
Challenges we faced
- Conversation continuity: Keeping visual context across multi-turn edits while routing to the right generation model and skill
- Production quality: Balancing speed, resolution, and consistency for e-commerce and marketing use cases
- Unified billing: Atomic credit deduction across async AI tasks, subscriptions, and one-time purchases
What's next
More MiniApps, stronger character consistency, team workspaces, and deeper API access for developers.
Built With
- better-auth
- cloudflare
- drizzle-orm
- gemini
- openrouter
- postgresql
- react
- replicate
- stripe
- tailwindcss
- tanstack
- typescript
- vercel-ai-sdk
- vite
