Goal
This week marks the first week of the summer-long project. For the first 2 weeks of this internship, I became familiar with making a dashboard using WDW products.
Moving forward, I have 2 goals for this week:
- Investigate my SkyGrade dashboard to figure out if the data being shown is accurate and not hallucinated by AI
- Continue brainstorming for the summer-long project
Workflow
When I showcased my SkyGrade sunset-quality dashboard, I received constructive criticism on how to improve the product. It was pointed out to me that all the parameters used to calculate a sunset score were being fetched from the point directly above the location, rather than towards the west, where the sun sets. Another point raised with me was the elevation factor. Before implementing any changes, I handed off my project to a team member to determine whether the data my dashboard pulls is accurate and up to date.
I’m looking forward to seeing what they uncover!
I had some conversations with Claude Code about the data-verticality issue, and I was surprised to see them draw this diagram, which helped me visualize the changes needed. Image 1 shows the diagram. I decided to let Claude attempt the major fix and some other minor changes. Image 2 shows the changes in place. There are a few tweaks:
- 2 parameters for the sunset score have been changed to take into account the Western horizon. This is why you can see long, straight areas of low sunset scores towards the west.
- There was also some potential data hallucination, especially over the open ocean. To address this, I decided to show the score only over land, not over the open ocean. While I’m not sure if this was the final fix, it did help a bit.
Until I hear back on the dashboard’s actual data integrity, I’m going to leave it as is for now and avoid making any other major changes.
For the rest of the week, I continued brainstorming for the summer project. My plan so far is to make a dashboard for air turbulence when flying. I propose two versions of this dashboard: one for pilots and one for fliers. The passengers’ dashboard will be simple, easy to read and understand, and reassuring. On the other hand, the pilot’s version of the dashboard would be much more technical and contain the specific information they need. Images 3 and 4 show a first-draft mockup of what the dashboards could look like, thanks to Claude Design.
Prompts Used
All my prompting this week was conversational, especially with Claude Code. I’ve found this to be a quick and easy way to get specific answers about the project code, while using an AI-generated prompt tends to work better for large requests.
What Worked
- Claude Code can be used like an LLM, especially for debugging or general questions.
- Additionally, Claude Code did a good job of visualizing its proposed changes. I found this very useful.
What Didn’t Work
- The main con this week was that my dashboard was slow. This is because AI tends to overdo things, making future debugging harder.
Lessons Learned
For my summer-long project, I’m going to try to establish strict rules for the AIs to follow. This should make things easier in the long run.
Images/Video
Image 1 (Click to enlarge)
Image 2 (Click to enlarge)
Image 3 (Click to enlarge)
Image 4 (Click to enlarge)