Inspiration
I took inspiration from my own experience. Every time I plan for a trip, I find that there are too many resources to look into, too many options to choose from. How can you be sure that the plan you are building will fit your needs?
What it does
Jidi tries to solve this problem by understanding exactly your needs to create a personalized experience.
How we built it
Jidi is built upon a standard frontend (React) / backend (Python FastAPI) architecture. The novelty is that interactions with the LLM are not done through a chat interface. Instead, Jidi interacts with the user through a Q&A system with choices. This avoids many issues such as prompt injections.
Challenges we ran into
As always, prompt tuning is not an easy task. The smallest changes can have significant impact on the quality of the generations. Unfortunately, I did not have time to try out monitoring/evaluation frameworks like W&B or Phospho.
Accomplishments that we're proud of
Building an end-to-end AI app in a week-end.
What we learned
Mistral Large is fast and impressively accurate. Also JSON formatting is less an issue than with other LLMs.
What's next for Jidi
Jidi needs to be backed up with factual data. The next step is to gather trusted, high quality data for each city supported and use it to ground the planning generation.
Log in or sign up for Devpost to join the conversation.