DEV Community

local ai
local ai

Posted on

I built an AI tool that turns survey data into research papers — here's the architecture

I built an AI tool that turns survey data into research papers — here's the architecture

Data2Paper Cover

Hey DEV community! I'm a solo founder building AI tools for researchers. My latest product is Data2Paper — it takes raw survey/questionnaire export data and produces complete research paper drafts.

The problem

Researchers collect survey data → export CSV → spend weeks turning it into a paper.

The manual workflow looks like this:

  • Clean the exported data (fix encoding, remove junk rows, identify the actual response sheet)
  • Recode variables and set up analysis frameworks
  • Run statistical tests in SPSS/R/Python
  • Build tables and charts
  • Write methodology, results, and discussion sections
  • Format everything into a deliverable document

Data2Paper compresses that entire workflow into a single pipeline.

Architecture overview 

┌─────────────┐
│  Upload      │  CSV / XLSX / XLS
│  (Survey     │  from any questionnaire platform
│   Export)    │
└──────┬──────┘
       │
       ▼
┌─────────────┐
│  Data        │  Identify response sheet vs summary
│  Intake      │  Parse machine headers (Q1, SC2...)
│              │  Detect variable types
└──────┬──────┘
       │
       ▼
┌─────────────┐
│  Analysis    │  Python execution chain
│  Engine      │  Statistical tests based on variable types
│              │  Generate charts & tables
└──────┬──────┘
       │
       ▼
┌─────────────┐
│  Paper       │  Multi-language (7 languages)
│  Generation  │  Full academic structure
│              │  Claude API
└──────┬──────┘
       │
       ▼
┌─────────────┐
│  Export      │  PDF / Word / LaTeX / ZIP
│  & Delivery  │
└─────────────┘
Enter fullscreen mode Exit fullscreen mode

Key technical decisions

Why Python execution instead of LLM-generated stats?

Language models can hallucinate numbers. For a research tool, that's unacceptable. The analysis engine runs actual Python code to compute statistics — correlation, regression, chi-square, ANOVA, etc. The LLM interprets the results, but doesn't generate them.

Why survey-specific, not generic?

Generic "data to text" tools don't understand that row 1 might be a machine header, that columns might represent Likert scales, or that the first sheet might be a summary rather than raw data. By focusing specifically on survey exports, the system handles these patterns reliably.

Why multi-language from day one?

Research is global. A tool that only outputs English misses a huge segment of users — Chinese grad students, European consulting teams, Japanese research groups. Supporting 7 languages in the generation pipeline (not as translation) was a deliberate product decision.

Tech stack

  • Frontend/Backend: Next.js on Vercel
  • AI: Claude API
  • Analysis: Python execution chain
  • Payments: Stripe
  • Export: PDF, DOCX, LaTeX rendering

Try it

If you work with survey data or know someone in academia who does: datatopaper.com

I'd love feedback from the DEV community, especially around the analysis pipeline design and the multi-language generation approach. Drop a comment or reach out!

Top comments (0)