Doujin Studio

AI-assisted video editor with chat-controlled timeline edits, instant previews, and versioned autosave

Comment

Inspiration

Video editing tools are powerful, but getting to a first cut is still a lot of repetitive clicking, scrubbing, and waiting on uploads. We wanted “edit video like text”: describe the change, see it applied instantly, and keep every step auditable and undoable.

What it does

Doujin Studio is a browser-based video editor where you can upload a clip, get an instant local preview, and edit a versioned timeline. You can also use an AI chat panel to request edits (like trims/reorders); the assistant streams a response and applies validated commands to create a new saved timeline version.

How we built it

  • Monorepo with pnpm + Turborepo
  • Web editor: Next.js (React) + Tailwind + Radix UI
  • API: Hono on Cloudflare Workers
  • Auth: Better Auth (email/password sessions)
  • Data: Drizzle ORM + Cloudflare D1
  • Media: Cloudflare R2 using short-lived pre-signed upload URLs + secure file streaming (range support)
  • Shared contracts: Zod schemas/types shared between web + API
  • AI: Google Gemini via the AI SDK with streaming + constrained tool calls that mutate the timeline safely
  • Tests/runbooks: Vitest + deterministic AI test mode + end-to-end validation guide

Challenges we ran into

  • Making uploads feel instant while still being reliable (local preview + background R2 upload + seamless swap)
  • Browser-to-R2 CORS and credential/session edge cases (especially hostname mismatches)
  • Designing timeline autosave + optimistic locking + version history so edits (human or AI) don’t corrupt state
  • Keeping AI edits safe: strict validation, bounded tool calls/commands, and graceful failure handling on an edge runtime

Accomplishments that we're proud of

  • End-to-end flow: sign up → upload → instant preview → timeline CRUD + autosave + versioning
  • AI chat that actually performs timeline edits through structured commands (not just “suggestions”)
  • Shared contracts across frontend/backend to prevent schema drift
  • Practical safety rails: rate limits, tool bounds, optimistic locking, and automated tests

What we learned

  • Streaming AI + tool execution is only useful when backed by strong contracts and validation
  • Pre-signed uploads and correct CORS policies are the difference between “works locally” and “works for users”
  • Versioned timelines make AI-driven editing explainable, debuggable, and reversible

What's next for Doujin Studio

  • Expand the command set (multi-clip operations, smarter trims, bulk edits) and improve edit explanations
  • Add background processing for heavier tasks (analysis/transcription/thumbnails) with queued workflows
  • Turn the rendering/export path into a first-class pipeline (building on the existing Remotion app)
  • Collaboration and sharing (invite teammates, comment on versions, share view-only cuts)

Built With

Share this project:

Updates