Cineserye

An AI filmmaking workspace that turns an idea into a research-backed, creator-reviewed scene plan, with a separate local image and video production workflow.

Comment

Inspiration

A filmmaker's idea can get lost between research tabs, story notes, shot lists and generated clips. We wanted a workspace where a creator can see what informed an AI suggestion and decide what becomes part of the film.

What it does

Cineserye connects ideas, scenes and shots on a visual canvas. In the recorded local researched-planning workflow, the creator approves a Parallel Search request, reviews retained sources, then separately approves a Gemini planning request. The proposed scenes, shots and claimed source influences are visible before the creator accepts them. Acceptance can be undone, redone and preserved across reloads.

The native production workflow supports reference-image generation, look approval and individual video takes. Our demo follows one separate three-second motion study through generation, playback, download and acceptance. It does not present the original 60-second plan as a fully generated movie.

How we built it

The application uses React, TypeScript, React Flow and runtime-validated cinema documents. Google's TypeScript Agent Development Kit and GenAI SDK run structured planning with Gemini on Vertex AI. A server-side Parallel Search integration retrieves the research used in the reviewed planning workflow.

AI output enters through a validated command boundary rather than writing directly to browser state. Paid requests have operation identities and explicit admission checks; uncertain outcomes retain their reservations instead of silently triggering another paid request. Native image and video production is distinct from the narrower Cloudflare research/planning target.

Codex helped develop the application, write contract tests, investigate errors, audit release materials and prepare the demo. Independent critic reviews were used to challenge completion claims and identify gaps. Codex is a development assistant, not the submitted application's runtime model.

Challenges we ran into

A citation can be structurally valid without supporting the creative detail attributed to it. We added source and influence review so a creator can distinguish sourced observations from creative additions. We also separated recovery from dispatch, and worked to make local-versus-hosted capabilities clear.

Accomplishments

The demo shows real Parallel research and a real Vertex-generated three-scene, nine-shot plan, followed by creator review, acceptance, undo/redo and reload persistence. A separate generated reference image and three-second video demonstrate the local production path. The clip was downloaded and verified against the generated output.

What we learned

An agentic filmmaking tool needs more than a persuasive result. Creators need inspectable inputs, clear approvals, reversible edits and honest boundaries around what has actually completed.

Scope and demo disclosure

The 2:45 demo is an edited sequence of genuine local application captures, with shortened generation waits disclosed. It includes Google Orus AI narration, explanatory English captions and the actual generated short clip at natural speed. It is not a continuous screen recording or proof of hosted image/video generation. Uploaded-photo likeness continuity is not guaranteed. Native Final Cut work exists separately, but this demo does not claim a live generated full-length assembled film.

What's next

Improve the research-to-shot review experience, broaden tested project sizes and make more of the native production workflow safely available to reviewers.

We also want Cineserye to become the video-production tool in a developer's agent stack. Planned integrations with Codex, Claude Code and other coding agents through MCP and a documented API would let developers turn a product brief and approved app captures into storyboards, narrated product demos, launch videos and social clips. The goal is to move from building an app to showing and promoting it in one connected workflow, while keeping creative review and generation-cost approvals in the user's hands.

These are future improvements, not claims about the current hosted feature set. The proposed coding-agent integrations are not part of the submitted application's current runtime.

Built With

Share this project:

Updates

posted an update —

Reviewer access note: The reviewer-code field currently has no visible border. Click the “Private reviewer code” label to focus it, paste the code from the private judging attachment (reviewer-instructions.txt), then select “Enter reviewer workspace.” Typed characters are masked. No post-deadline changes have been made to the submitted app.

Log in or sign up for Devpost to join the conversation.

Submission history