I was a key contributor in developing Designstripe’s social post generation product. I worked with the CTO to plan and implement an ideal pipeline to take a Brand DNA and a topic and generate at least 90 brand-appropriate graphic designs geared towards social media formats with copywriting crafted by LLMs - all in under 5-10 seconds.
I was responsible for developing the backend and frontend for this system, while the CTO built out a robust system for fulfilling LLM requests using multiple providers to avoid capacity or downtime issues for any given provider. This was a common problem in the early days of LLM-driven workflows.
Together we were able to hit our target, by utilized vector embeddings to find relevant imagery for the graphic designs and progressively generating the graphic designs as soon as enough copy was available for the first “page” of data (about 6 designs), we were often showing content between 2-5 seconds.
Once content was generated, a user could click on any design to customize it. Including recoloring, using AI to improve text or generate new ones, use their own images, or enter an advanced mode closer to traditional graphic editors where the design was split into layers that could be manipulated freely.
Once a user was happy with a design, we built a publishing system where the user could link their social media account for a target platform and have the graphic design automatically posted there on their behalf. This product allowed for a tight loop for brands to enter a topic of their choice, find a great design, and publish it in one session.