Dreamschemas

Converts CSV files into a Supabase Postgres database schema with AI schema generation and seamless Supabase project creation and import.

Comment

Inspiration

The entire world can now vibe-code fully functional applications with databases thanks to Supabase and tools like Bolt.new. But here’s the catch: setting up a real database for your project is still a massive roadblock, especially for newcomers. If your business lives in spreadsheets, getting that data into a legit, production-grade Postgres schema is a nightmare. Dreamschemas was born to bridge that gap, making it dead simple to turn your business’s messy CSVs into a real, usable database foundation.

What it does

Dreamschemas lets you upload multiple CSVs at once, automatically pre-validates them (with clear errors, warnings, and info), and then sends them to AI to generate a custom schema complete with relationships and row-level security. You can export the schema to .sql migration files and other formats, or go further: authenticate to your Supabase org with OAuth2, pick your project, and have Dreamschemas auto-create the initial schema for you. But it doesn’t stop there the schema is then used by AI to generate a custom edge function, letting you upload your CSVs to Supabase S3 object storage and import them in batches straight from that edge function. No manual setup, no guesswork, just a direct path from spreadsheet to scalable, secure database.

How we built it

We used a process that starts with an "AIR" doc ("AI Requirements" like a PRD for LLMs), created in markdown in Cursor as the first file in a repo. We then flush out a project knowledge doc to be used in tools like Bolt. We use the Vercel Next.js Supabase boilerplate in a repo and sync Bolt with that using Github Sync. We then had Bolt build the landing page. Next we pulled that down locally and did the first pass with Claude Code parallel agents. Next we edited in Cursor with Sonnet 4. Then back to Bolt for visual edits, rinse and repeat.

Challenges we ran into

  • Making AI actually understand messy, real-world CSVs (CSVs are incredibly non-standard)
  • Translating spreadsheet columns into tables with related columns
  • Getting AI to get column constraints correct
  • Getting AI to create a good Edge Function script
  • Running Edge Functions in batches for large files
  • Getting past RLS and permissions issues for seeding data

Accomplishments that we’re proud of

  • 90% of the flow works perfectly
  • The schema generation and database table creation are fantastic
  • Our workflow for using Supabase migrations, edge functions, object storage is smart

What we learned

  • You have to use a mix of lots of AI tools for something this complex
  • Seeding data is SO hard
  • If we pull this off it will be a massively populate open source tool
  • Bolt is far more mature than we recently assumed

What’s next for Dreamschema

  • Get seeding more reliable
  • Release to the open source community
  • More postgres database targets beyond Supabase
  • Even faster, even prettier, even less manual work
  • Look for open-source contributors!

Built With

Share this project:

Updates