Marketing Studio

Inspiration

Marketing Studio started with a problem I experienced while creating my own videos: generating content was only one part of the work. I still had to choose the right footage, assemble scenes, refine the edit, and render the final video.

When I applied for a Software Engineer position at 8x, their assignment to build a Higgsfield-inspired product gave me an opportunity to work on that problem.

I wanted to build a creative workflow where AI helps produce the content while the creator keeps control of the story. That became the foundation of Marketing Studio.

What it does

Marketing Studio brings several creative workflows into one application:

Avatars: Create and save an identity to reuse across images, video templates, and effects.
Visual tools: Work with image generation, image editing, and identity-based video effects.
Audio: Generate narration, music, and sound effects.
Documentaries: Turn a topic into an editable script, review footage choices for each scene, and select the shots that fit.
Custom editor: Arrange clips, adjust timing, and combine visuals with narration and other audio.

The documentary workflow offers up to three stock-footage candidates per scene, depending on availability. Selected footage and narration load into the same project’s editing timeline.

A cloud rendering pipeline is integrated to move the final encoding work off the creator’s device. Live export verification is still in progress.

The application also includes Google sign-in, saved projects, usage-based plan limits, and a payment gateway integration.

How we built it

I built the frontend with React and TypeScript and the backend with Express. The frontend runs on Vercel, while the API runs on AWS EC2.

The backend coordinates specialized services:

RunPod GPU workers host Krea and Qwen image models.
Modal provides the infrastructure for LTX video generation.
RunPod CPU workers handle FaceFusion processing and the FFmpeg rendering pipeline.
OpenRouter and ai33pro provide additional hosted model and audio capabilities.
Pexels supplies footage candidates for documentary scenes.
Cloudflare R2 stores media, while Supabase stores application data.

Generation runs asynchronously. The API creates a job and returns promptly, then the frontend tracks its progress while a worker processes it.

I also integrated Google OAuth, plan quotas, and SwichNow hosted checkout to build the foundations of a usable product.

Challenges we ran into

One challenge was making different models and services behave consistently. Each provider has its own inputs, response formats, processing times, and failure conditions. I had to bring them together behind a shared job-tracking workflow.

Infrastructure introduced another set of problems: cold starts, model dependencies, storage configuration, and worker availability. During one incident, RunPod reported both ready and throttled workers, but the application treated any throttling as a complete outage. I corrected the availability check so usable workers could accept work while retaining protection when none were available.

The documentary editor also required careful coordination. Scene selections, narration, media ownership, and timeline changes needed to stay connected to the same project.

Finally, I had to distinguish between integrating a capability and verifying it end to end. Cloud export remains an area where the implementation needs final live validation.

Accomplishments that we're proud of

I’m proud of turning an assignment into a deployed product built around a problem I understood personally.

The strongest accomplishment is the documentary workflow: users can review the script, choose footage scene by scene, and continue editing their selections in one project.

I’m also proud of the engineering beneath the interface: asynchronous processing, specialized workers, persistent projects, media storage, authentication, usage controls, and payment integration.

Building across these layers taught me how much care it takes to make individual features work together as a product.

What we learned

I learned that model quality is only one part of an AI product. Users also need understandable progress, reliable job tracking, useful errors, and control over the result.

Open-source models offer flexibility, but they also bring responsibility for deployment, dependencies, storage, and recovery.

I learned to choose infrastructure around the workload. Image generation, face processing, and video encoding have different requirements, so they should not all compete for the same resources.

Most importantly, I learned to treat operational failures as product problems. A worker-status check can determine whether someone finishes their creative work or gets stuck.

What's next for Marketing Studio

My immediate priority is to complete live cloud-export verification and test the entire documentary workflow from topic to downloadable video.

Next, I want to improve worker recovery, make progress and error messages clearer, and test the experience with creators making real projects.

Their feedback will guide improvements to scene selection, editing, and output quality. I also want to measure actual processing costs and completion times before making specific affordability or performance claims.