ChatPaperAi

• 

  ChatPaper landing page — Upload → Analyze → Read → Export flow

Inspiration

Every day, millions of meaningful conversations happen on KakaoTalk — study groups debating research ideas, teams solving complex problems, friends exchanging deep insights. Yet these conversations disappear, unrecognized as the intellectual work they truly are.

We asked: what if a chat log could become a citable academic paper?

What We Built

chat-paper-platform is a Korean-first SaaS platform that transforms KakaoTalk and AI chat exports into fully structured academic papers — complete with abstract, methodology, results, and references.

The pipeline works in 6 stages:

1. Parse — Extract speakers, timestamps, and message threads from raw .txt exports
2. Detect — Identify language and conversation structure
3. Analyze — Topic clustering and sentiment scoring using NLP
4. Generate — Feed structured data into a 6-step LLM prompt pipeline (GPT-4o)
5. Format — Render as a proper academic document
6. Export — Download as PDF or DOCX

The sentiment score per speaker is modeled as:

$$S_i = \frac{1}{N} \sum_{j=1}^{N} \text{sentiment}(m_{ij}) \in [-1, 1]$$

How We Built It

• Frontend: Next.js 14 App Router + TailwindCSS + shadcn/ui
• Backend: Next.js API Routes + Prisma ORM + PostgreSQL (Neon)
• AI: OpenAI GPT-4o with a 6-stage prompt pipeline
• NLP: Custom KakaoTalk parser, language detection, topic clustering
• Export: PDF and DOCX generation with multilingual font support
• Privacy: Local anonymization — names and contacts masked before any data leaves the device

Challenges

• KakaoTalk format parsing is undocumented and changes across app versions — building a robust parser required handling dozens of edge cases
• Academic tone generation in Korean required careful prompt engineering to avoid machine-translation artifacts
• Privacy by design — ensuring personal data never hits the server unmasked was a non-trivial architectural constraint

What We Learned

Conversations have latent academic structure — argument, evidence, counter-argument, conclusion. LLMs are surprisingly good at surfacing it, given the right scaffolding.

The biggest insight: the hardest part wasn't the AI. It was the parser.

Built With

• docker
• docx
• next.js-14
• nextauth.js
• node.js
• openai-gpt-4o
• pdf-generation
• postgresql-(neon)
• prisma-orm
• redis-(upstash)
• shadcn/ui
• tailwindcss
• typescript
• vercel
• zerve