-
launch vehicle
Public Loom: https://www.loom.com/share/e7aa1e765f6244128de1e61d8e7cf090
Inspiration
while we love the content of grokipedia, it's hard to read an article in it for more than five minutes. three main reasons why an uplift to the user experience will be a win-win situation for all:
- grokipedia becomes widely accessible to people, creating a dense and unbiased learning experience compared to traditional learning experiences
- with more people on the platform, xAI will benefit with more data for improving articles across various dimensions
- eventually the set of facts for many topics will become too large for most readers to consume, so different people will need to be shown different projections of the data
What it does
- enhances articles with multimodal information
- has a personalization engine based on GEPA (https://arxiv.org/abs/2507.19457) to tune articles based on user liking and preference
- swiping based UI to create data for the personalization engine
- brainrot scrolling UI for the general population
How we built it
- hosted on https://yourgrokipedia.com/
- nextjs app + postgres db deployed on an autoscaling infra on GCP
- google oauth
- xai API for the GEPA algorithm and content personalization and rewriting
- xai image generation API for synthetic images
- elevenlabs with custom audio for the brainrot + SSE streaming responses
- added caching support for all content for faster reloading
- used cursor
Challenges we ran into
- connecting everything together and actually deploying
- running GEPA properly as typescript support was just added recently and is fragile
- karaoke style subtitle sync for the brainrot UI (didn't really work)
- ensuring that we don't go into an quick autoreloading cycle and consume all API credits
Accomplishments that we're proud of
- quick and rapid prototyping and deploying quickly, learned a lot of new tricks to speed up things
- GEPA is fragile sometimes, but outputs actually look good
What we learned
- how grokipedia is built and works
- audio APIs, SSE
What's next for your grokipedia dot com
- combine synthetic images with real ones obtained from citations - text is only a part of whole truth, some truth is contained in images and can't be really converted without loss of signal
- improve the GEPA optimization process
- add different synthetic videos to the scrolling UI
Built With
- agent-optimization
- elevenlabs
- evolutionary-algorithms
- gcp
- gepa
- nextjs
- postgresql
- python
- xai
