BUILDING MONROE - Part III
Last modified: 2024-06-27
This is part 3 of a series on building MONROE. You can find part 1 here and part 2 here.
Key features
Adding the ability to watch shows together
I knew from the outset that one of the core features I wanted to add was
the ability to a watch a show with your friends. I wanted to make it seemless to
not only share what you've been watching regardless of streaming platform but
also to know where your friends are in a show so you can catch up and watch it
together.
I thought this would be pretty straightforward -- just add a "watch together" button, right? Turns out it was much more complex than I was expecting.
I started by creating a model for a UserConnection that would store the
userId of one person and the userId of the person being invited to watch. So I
added the model of a UserGroup. This would connect the UserShow (the model of an
instance of a user watching the show) to a group. This allowed for multiple
people to be in one group and watch the show together. So then I just needed to
add the ability to add people to the group. Simple enough, right? But you don't
want to just allow anyone to add anyone else to a group? What if you don't know
the person inviting you? Or just don't want to join the group?
So you need the concept of a UserInvite as a limbo state between being invited and joining the group. And then you need to add the ability to accept or decline the invite. But you also don't want to have to accept every invite every time you want to watch with a group. Think of a groupchat -- sometimes its nice to just be spontaneously be added by someone you know and then leave the group if you want. So you also need a model of a "UserConnection" to indicate that the invitees are connected and can watch together immediately. So just to allow this one feature, I needed to add three additional models and manage the state of each of them.
AI Chat Engine w/ Personalization
There's a whole world of people building "thin wrappers" around ChatGPT -- so I thought, hey - how about one more?
Relevant News -- Bulk Categorization using AI
I didn't just use AI for the chat engine to get recommendations. One of
the features I wanted to build was to pull in news articles about the shows
you've been watching. This seems simple and the scraping part of it is. Just
used BeautifulSoup to scrape the articles and then store them in a Postgres. The
hard part was categorizing them accurately and in a scalable fashion. Even if it
was obvious that an article was about, say, the show Succession, I needed to be
able to tie that back to the exact show id in my database. Now, I could do that
manually for a couple but for all of them, that was not going to happen.
So I first pulled all shows from my database along with their show ids
and descriptions and fed that to the OpenAI database as part of a system prompt.
For context, Monroe has about 176k shows in its database so this was a
substantial amount of data. I then fed each article into the OpenAI API and
asked it to output a JSON object with the article_url and show_id. I took this
output and stored it in Postgres. Altogether, it took about $30 to process but
saved me days of manual work and I ended up with around 10k categorized articles
to seed the news feed.
Internal Tools
What I did Too Soon
Dark Mode Exploring a new idea is really exciting because it's a blank canvas
on which you can exercise your creativity. That being said, prioritization is still important and there are some things that I wish I had prioritized differently. Specifically, I'm thinking about adding a Dark Mode switch. I'll start by saying that using the app at night with a darker theme is much easier on the eyes. Now that I have it, I do really like it. But I wouldn't have started it so early for two reasons: 1) it's not a core feature that affects the core user experience or materially addresses the problem the product is trying to solve and 2) it slows down the development of new features as now everything has to have both a light and dark theme. Yes, eventually I would have had to do this but it would be much easier to do it all at once rather than piecemeal.
Episode Specific Conversations
A second feature that I wish I had waited on was the chat features. It's a
feature that isn't useful in isolation. So you have a cold-start problem as no
one wants to start the conversation and the action-reward cycle is essentially
non-existent if there are not others engaging in the discussion with you. Unless
you have high activity, it's going to feel like a ghost town. So why did I build
this so early? Well, put simply, it came out of a discussion on Product Hunt and
I thought it would be a fun challenge to explore.
For something like a movie, which exists in a binary-state -- you've either seen it or you haven't. In that case, it's pretty easy to design a chat feature that allows you to discuss the movie as a whole. You can organize the chat by topic threads to discuss themes, events, reveals, theories, etc but the only piece of information you need going in is whether or not the user has seen it.
Television is different. It's story is serialized and can be told over dozens (or even hundreds) of installments taking place over years both in-world and in real-time. The show is stopped at discrete invtervals, with the viewers understanding of the show changing depending on where they have stopped. Designing a chat feature in this environment is much more complex because you need to know where the user is in the show to avoid spoilers.
As a first step, I added the ability to start a chat on a specific episode basis. If you click on an episode, a popover window opens with the current discussion. If you have seen the episode, it will show the current discussion. If you haven't, it will show a warning that there may be spoilers ahead. This required me to model the concept of a conversation that can be associated with the show and episode and then a message concept that can be attached to a conversation. This was a great way to expand out the data model but on reflection, I can already see ways that its limited.
Next Steps
* I do feel compelled to say that we did launch a product that made a significant impact and which (from what I've heard from people are still there) is growing at a steady rate. This may seem petty to say but I do want to shout out everyone who worked on this product and say that we were able to build something great despite everything.