Blog cover image

Leveraging AI to create personalized video at scale

BySarimDec 19, 2023

Context

As 2023 comes to a close, Graphite wanted to celebrate GitHub users for their contributions throughout the year. The goal was to end the year with a gift for developers to reminisce, reflect, and feel inspired for the new year.

As the creators of GitHub Wrapped, a project we built in 2021 and scaled to 10k users, our team at Rubric was perfectly positioned to take this on.

However, 2023 was unlike any other year. 2023 was the year LLMs became generally available.

Compared to 2021, it felt like the realm of opportunities had opened wide for us and we wanted to push past static images and templated storylines, as we had done so before. Instead, we wanted to create something truly personalized, completely unique to the end user. We also wanted this to be immersive. A Year in Code was born — personalized, AI-generated video.

It’s no surprise that we ended up leveraging LangChain to build this. LangChain’s out of the box helper functions helped us get to production in days, rather than weeks.

Tech Stack

Architecture

Overview

Let’s summarize the architecture in a simple diagram.

Flowchart showing the architecture of Year in Code. GitHub API, gpt-4-turbo, LangChain and Zod, Remotion, AWS Lambda, to video.

We begin by authenticating a GitHub user using Supabase auth. Once authenticated, we fetch user-specific data from the GitHub GraphQL API, and store it in our PostgreSQL database hosted on Supabase. Supabase offers an out of the box API with Row Level Security (RLS) which streamlines reads/writes to the database.

At this point, we pass user stats to the LLM (gpt-4-turbo) using LangChain. Leveraging prompt engineering, function-calling & Zod schema validation, we are able to generate structured output called the video_manifest. Think of this as the script of the video.

This manifest is passed to a Remotion player which allows easy embeds of Remotion videos in React apps at runtime. The manifest maps over a set of a React components.

At this point, the user is able to play the video in the client and also share their URL with their friends. Next.js 13 server rendering patterns make this seamless for the end user. Additionally, the user is able to download an .mp4 file for easy sharing by rendering the video in the cloud using AWS lambda and storing the video in an S3 storage bucket.

Let’s explore this in greater detail.

Fetching stats

When you log into the app with GitHub, we fetch some of your stats right away. These include

  • your most-written languages,
  • repositories you’ve contributed to,
  • stars you’ve given and received, and
  • your newest friends.

We also fetch your total commits, pull requests, and opened issues. Check the type below to get a sense of the data we fetch. We wanted to be cognizant of scope here so we ask for the most necessary permissions, excluding any access to code. The project is also fully open source to reinforce trust with the end user.

Code snippet of the interface for stats, with username, year, etc.

Generating the manifest

We then pass these stats to OpenAI’s gpt-4-turbo, along with a prompt on how to format its response. Here’s the prompt:

Code snippet of the prompt template: you are GitHub Video Maker, etc.

Given user stats, the AI generates a video_manifest which is similar to a script for the video. The manifest tells a unique story in 12 sequences (as defined in the prompt). Assuming each sequence lasts 5 seconds, this results in a 60 second video consistently.

Here we ran into a challenging problem: do we give the AI complete creative freedom or do we template as guardrails for the AI?

After running some experiments, we quickly realized that in the given timeframe, we couldn’t generate high quality video by giving AI the complete creative freedom. While the output was decent and could have been improved, it wasn’t good enough to have that nostalgic moment, especially in the engineering time that we had.

So instead we struck middle ground by creating a bank of “scenes” and parametrized them as much as possible. This allowed the AI to pick from a bank of scenes, pass user-specific data & generate unique text for each scene, resulting in a unique sequence of personalized frames.

This was possible using OpenAI’s Function Calling which enabled the AI to output parsable text, conforming to a Zod schema. The schema uses a Zod discriminated union (not the name of a rockband) to distinguish scenes:

Code snippet of the video schema in Zod: array of objects with text and animation, etc.

See the full schema here.

Let’s look at a sample output video manifest.

Code snippet of the video manifest: text, animation, etc.

Each entry (scene) in the manifest is an object that has a text field and an animation field. The text is unique for each scene and so is the order of the scenes, whereas, the animation for each scene is picked form a bank of pre-built components.

Playing the video

Now the fun part: playing the actual video. This part was challenging, because we’re quite literally letting an AI direct a video we’ll trim together. From that director’s cut, we map scenes to React components, which Remotion takes to generate a video. Take a look:

Code snippet of the video scene mapping in React

Here, the from prop determines the first frame when this scene will appear.

To generate 3D objects, we leveraged Three.js. For example, to mould this wormhole effect from a flat galaxy image, we pushed Three’s TubeGeometry to its limits with high polygon count and low radius.

Flat purple nebula image with an arrow pointing to the version mapped to a 3D tunnel shape

Now, we want this experience to scale by being as lightweight as possible. By saving the video_manifest, instead of the actual video, we trim the bulk of the project’s bandwidth and storage by 100x. Another benefit of this approach is that the video is actually interactive.

Rendering the video

Since we map over a manifest in the client using React components, to download the video as .mp4, we have to render the video first. This is achieved using Remotion lambda leveraging 10,000 concurrent AWS Lambda instances and storing the file in an S3 bucket. Each user only has to render their video once, after which we store their download URL in Supabase for subsequent downloads.

This step is the most expensive in the entire process and we intentionally added some friction to this step so that only the users that care the most about sharing their video end up executing this step.

Conclusion

This project makes use of all the latest tech. Server-side rendering, an open-source database, LLMs, 3D, generative video. These sound like buzzwords but each is used very intentionally in this project. We hope it inspires you to build something new in 2024!

Ready for takeoff? Give Year in code a try. Translate your keystrokes into stardust. Find solace in your retrospection, let others join you in your journey, and connect with starfarers alike.

Your chronicle awaits.

If you have feedback on this post, please reach out to us at hello@rubriclabs.com

Rubric Labs

Let's make something amazing, together.

Get in touch