Show HN: NailGenie – Edit nail designs conversationally with AI
nailgenie.org
I built NailGenie (https://nailgenie.org) to solve the "that's not what I meant" problem in nail design. It's an AI platform that lets you iteratively edit nail art through simple conversation rather than static generation.
THE TECHNICAL CHALLENGE
The core challenge was building a system that could understand contextual, incremental editing commands for a specific visual domain. Most generative AI solutions focus on one-shot generation, not a continuing dialogue about the same image.
We solved this by:
1. Fine-tuning Gemini on a dataset of nail designs with paired editing instructions
2. Building a stateful context management system to track editing history
3. Creating a visual diffing algorithm that preserves nail boundaries during edits
4. Implementing an instruction parser that handles ambiguous editing requests
The backend reaches ~98% instruction comprehension on our test set and produces edits in ~2.7 seconds on average.
TECH STACK
- Frontend: Next.js App Router with TypeScript and React Server Components
- UI: Shadcn/UI + TailwindCSS (we chose these for rapid iteration)
- Backend: Supabase for authentication, storing edit history, and managing user credits
- Deployment: Vercel edge functions for low-latency API responses
- AI: Custom-tuned Gemini models with a multi-stage processing pipeline
DEVELOPMENT CHALLENGES AND LEARNINGS
The biggest challenges were:
1. Instruction ambiguity: "Make it more pink" means different things to different users. We implemented a clarification system that refines ambiguous requests.
2. Edge detection: Early versions struggled with nail boundaries. We built a specialized segmentation model to ensure edits only affected the nail area.
3. Performance: Initial processing was ~8s per edit. We optimized by parallelizing our pipeline and caching intermediate representations, cutting time by ~65%.
4. Cold starts: Edge function cold starts were killing the experience. We implemented background warmers and optimized model loading.
THE WHY AND WHAT'S NEXT
I'm not a nail expert, but I noticed my girlfriend spending hours browsing examples before salon visits, then being frustrated when the result didn't match her vision. The challenge of creating a system that bridges this communication gap became technically fascinating.
Current metrics: - ~450 users in closed beta
- Average session: 8.3 edits per design
- 82% completion rate (users reaching a final saved design)
FUTURE PLANS
- Open source our instruction parsing logic next month
- Add API access for nail salons to integrate directly
- Implement real-time collaborative editing
TRY IT YOURSELF
NailGenie is live with free starter credits. I'd appreciate any feedback, especially on:
- Instruction parsing accuracy
- Performance bottlenecks you experience
- UI/UX pain points