Launch Week Recap
We tried a couple of different things for launch week.
We first tried to table in Huang. We got a bunch of candy, brought our whiteboard out, and set up shop. Unfortunately, we picked a pretty bad day to go out tabling -- a Thursday. Compared to the same time other days, there was quite literally no one there, so we had a hard time getting people to show up. From there, we knew we had to pivot.
It was around lunchtime, so we thought it would be a great idea to go to a dining hall. We decided to go to Avanti in Suites, and we were able to get some foot traffic there.
Overall, we got some great remarks. We also get an average score around 4.
I liked that I could use words that typically would break a search engine.
It provided good analysis of the products it was suggesting.
I found the way of interaction was really intuitive and I could use natural language prompts as opposed to just searching for tags
I think it was an improvement by having natural language input and the ability to search things that supost isn't able to by natural langauge (like "furnishing my apartment")
Kenja allowed me to search for more general search queries, which is really helpful for when I am struggling with how to describe an item I have in my head. When I searched up dorm decoration without quite knowing what I was looking for, Kenja had much better responses. It provided us with tapestries, rugs, and other items that didn't necessarily come up on SUPost. It was able to identify the types of items I really wanted with a broader search then what SUPost was able to do. I noticed similar things when I searched up, "appliances." Kenja would pull up actual kitchen appliances which is what someone would likely be looking like in that scenario while SUPost would simply list apartments (which seemed to include appliances). It is more accurately able to identify what a user wants based on these broader terms. It makes the experience of using this app a lot easier.
Of course, there were also some areas to improve. In general, hallucinations were a big piece of feedback:
I think sometimes it would recommend very irrelevant things
It wasn't as focused on the search. SO it gave some options that weren't relevant to the query.
Some things I wanted to see on the Kenja search was the use of images or something that mimicked a traditional search more.
Another big piece of feedback was about the lack of images. This is more of just an interface delivery aspect, and not as much a hit on our algorithm.
To get more feedback on our search, we also decided to deploy our application to a larger EC2 instance and send it out to more peers. You can check out the demo here.