“WTF is that?” — the birth of an image recognition bot
WTFIT (WTF is that?) is a chat bot that describes the images you send it. It began as a little experiment with the Facebook Messenger platform, and quickly grew to 4000+ users and has processed over 16,000 images in just under 2 weeks. I will share my journey in building this bot as well as some learnings I have picked up along the way.
Inception
One thing that had bugged me for a long time was not knowing what something was. When someone told me something unfamiliar, I Googled it. When I heard a song on the radio, I put it through Soundhound or Shazam. What happens when I see something I don’t recognize but would want to know more about it? That’s how I came up with WTFIT.
This realization happened around the time of the F8 conference, where Facebook announced their Messenger Platform API. This allowed developers to build bots that connect to Facebook users through Messenger. This to me was a perfect platform to build WTFIT on top of, as it has almost a billion users and an interface that’s familiar to many. People would be able to access this service without going on to a website, downloading an app, or even signing up for anything.
Building the Bot
As soon as it was announced, I began building. I dropped everything I was doing and got a functional prototype running in a few days. To speed up the process, I submitted the prototype to Facebook to begin the review process early on, while I was doing further user testing with friends and refining it.
My approach to testing was quite different from a previous product I had worked on. Although I showed friends and asked them to play with it, none of them knew I had built it. That allowed me to observe usage as well as get good honest feedback. It also allowed me to refine the image recognition as well as some of the back-end data handling, so that it could handle potentially large volumes of requests.
Two weeks later, I received a notification telling me it was rejected. T’was a sad morning indeed. However, it turned out the problem was that they couldn’t find the page associated with it. That was strange, because all my friends could access it. I decided to resubmit, and noticed that I now had the option to select the page associated with the bot. In a weeks time, I started getting requests from test accounts that I assumed were from the Facebook team, and a day after that, it was approved. Wohooo!
Launch
So my bot was approved, now what? How do I tell people about it? Fortunately, I came across Matt Schlicht’s Facebook group called Bots. At the time it had about 2600 people all interested in chat bots, and decided it’d be great to see what they thought about it.
Initial feedback was very positive. It allowed me to further gauge the recognition performance for random image thrown at it, and the opportunity to test a sudden influx of traffic. I had to patch the server up a few times but fortunately there were no major issues nor any major downtime. The part I loved the most was people posting their results on the Facebook group, some amazing, some funny, and some where there was room for improvement.
The next step was to seek all the bot directories out there. The one that appeared at the top of my search was botlist.co. I tweeted the founders and Ben Tossell kindly added the bot to the directory. He also sent me an invite to Product Hunt, which I was really excited about.
It was 11 pm local time and I had the option to hunt the product. I did have an early meeting the next day but what the heck, I decided to do it.
Growth and Publicity
Listing on Product Hunt was when the traffic really hit. Upon hunting the bot, the server began to process requests more frequently by the minute. In less than 24 hours, it processed 2000+ images. The Product Hunt community was buzzing and there were lots of great comments left by them. Tech blogs across Spain, Germany, Japan, and US were talking about it, which brought even more international traffic. In the morning, I was contacted by WIRED writer Tim Moynihan who did a story on it. Safe to say, I was ecstatic.
Reflecting on this later, my decision to list the bot in the late hours of the evening was both a good and bad idea. 11 pm in New Zealand was approximately 4 am PST and 7 am EST, which was the start of the day in the US and helped it get noticed. It quickly bubbled up the list and snowballed into even more hits. That maintained a steady stream of people trying out the bot throughout the day. The downside was that I ended up staying up to 5 am to answer questions and keep the servers alive, and ended up getting about an hour of sleep. Oh well, it was well worth it!
Nerdy Stats
Product Hunt brought in the first big wave of traffic. Spikes of traffic over the last two weeks were brought in by local tech blogs around the world.
Over the last two weeks, there has been worldwide usage of the bot. I suppose with such a universal concept of taking a photo and sending it, there aren’t really any language or regional barriers!
It’s been great seeing what people have been sending in. These are some of the conversations the users have shared with us!
What’s Next?
WTFIT is currently a fun bot to throw all your photos and images at to see what it thinks. However, I believe it has the potential to become a useful image recognition tool. To further grow the bot, I have teamed up with Spark 64 Ltd (creators of UVLens) to build out the capabilities and scale up the platform. One day, you’ll be able to take a photo of that flower in your garden and find out the species, or that item of clothing to find where to buy something similar.
Exciting times ahead! Make sure you keep up with our progress on our Facebook page!
Special thanks to the Bots group, Product Hunt community, Spark 64 Ltd co-founders Daniel Xu and Richard McLean, and last but not least, friends and family Jan Tam, Johnson Cheuk, Herng Jeng Leong, Kejia Wang, Alyssa Ong, Hayden Do for testing the service in its early stages and your valuable feedback!
