A.I. Startup: Day 1
What does it look like for 3 non-PhDs to build a successful machine learning product? We’re not totally sure yet! Follow us as we document our journey building a web app that uses artificial intelligence without any formal training in the topic.
Who do you think you are?
We’re John, Darcy, and Kyle! Here’s a (terrible) photo of us celebrating after meeting with a lawyer.
Before we were co-founders we were co-workers building a git client at a 40 person software company. We worked hard and excitingly it payed off as our monthly active user count rose from 0 to 100,000 over the span of a year. We had built something that people loved, and it felt great!
After leaving that company we started solidifying the idea for a product that had been missing in the development process.
What’s the product?
A good problem to have when building an app is having too many users. Particularly for us, this was a problem because customers were flooding our support channels. Via email, Twitter, Facebook, our public Slack channel— customers were overwhelming us with their feedback.
While our customers’ suggestions should have influenced our product roadmap, the volume was just too high to read it all. So instead we had to settle for cherrypicked tweets and emails to get a sense of the overall sentiment.
If only we had a dedicated employee that read all the feedback and could tell us that 65% of users wanted Feature X and 70% were annoyed with Bug Z.
Enter Hashback — a feedback dashboard. Hashback aggregates feedback from Twitter, Facebook, surveys, and support calls, and applies text categorization and sentiment analysis to turn it into useful information. In a sense, Hashback is the dedicated employee that can tell you all about user’s thoughts towards Feature X and Bug Z.
How’s it going so far?
With 4 weeks behind us, the prototype is on the horizon and will likely be finished in the next two weeks. If you did the math and think 6 weeks is an insanely short amount of time for deep learning newbies to create two NLP-based neural networks then you’re right. We didn’t. Instead we started with IBM Watson for text classification and Google Cloud Platform for sentiment analysis. They’re both pretty cheap and allow us a working prototype while creating our own NNs.
When that sweet sweet build-the-neural-net day comes we’ll be prepared. Aside from the usual web app development agenda we’ve also been taking Andrew Ng’s deep learning specialization on Coursera. (It can’t be recommend enough if you’re looking to break into deep learning. Ng does an amazing job demystifying neural networks and you’ll be implementing one from scratch in a matter of weeks!)
What’s next?
The MVP is only a couple weeks away. Once it’s shipped we’ll start working on our custom neural networks and a demo page to encourage signups.
Conclusion
That’s us in a nutshell. The scrappy underdogs using cheap online courses as a propeller towards a machine learning product. We’re certainly not the first to do it, nonetheless we’re excited to share an inside look of the process from Day 1.
