Inspiration
Video editing turns a simple intention into many small interface actions. Timeline Studio lets creators request changes in plain language, review exactly what will change, and keep control of an editable project.
What it does
The competition edition adds an explicitly labeled simulated Alexa+ assistant to an existing browser video editor. Real local AI planning with Ollama and Qwen3 4B proposes aspect-ratio changes, simple video trims, clip reordering, music volume, fades and mute changes. A separate deterministic command mode is also available. Every proposed edit goes through a change-review panel before application. Creators can keep editing manually, undo the last unchanged agent transaction, save a portable project and render a video.
How we built it
The published codex/alexa-hackathon-2026 branch contains a separately loaded /competition experience in React and Vite. A local server sends bounded conversation and clip metadata to Ollama; source media stays in the browser. Project-specific structured output constrains the model to supported operations, current clip IDs and trim bounds. The existing editor validates the plan and checks a state token before applying it. An optional Bedrock adapter is included but was not live-tested or used. We enter the Alexa+ simulated-experience track; we do not claim a live Alexa+ connection or AWS Builder eligibility.
Challenges we ran into
The initial small-model response echoed project metadata instead of proposing edits. We fixed the prompt boundary and added structured JSON output constrained to the current project. Generated prose also claimed completion too early, so the interface shows a review instruction for proposals and reports success only after the actual editor receipt. State tokens reject delayed plans when manual work changes the project.
Accomplishments
On Apple M1, real Qwen3 4B inference completed an English request for a vertical ratio, a two-second first clip and music at 20%. A subsequent Chinese request reordered clips and muted music; undo restored the previous state. We reopened the saved seven-second project and fully decoded its 720×1280 H.264/AAC export without errors. Cancellation, stale-plan rejection and a 412×915 mobile layout were checked. Lint, typecheck and production build passed with existing warnings. No paid model API was used.
Significant hackathon updates
This existing editor gained the conversational competition assistant, local model adapter, bounded metadata context, project-specific structured output, guarded review/apply flow, cancellation and direct assistant copy in all thirteen supported interface languages during the hackathon period. The original editor remains available at the root URL.
What we learned and what's next
Metadata is not visual understanding, a proposal is not an applied edit, and a simulation must be labeled clearly. Current limitations include no video-pixel or speech understanding in the planner, no assistant voice-input UI and no conversation restoration after refresh. Next steps are richer supported editing tools and better performance while preserving transparent, reversible changes.
Built With
- javascript
- ollama
- react
Log in or sign up for Devpost to join the conversation.