Future Keys

keys from radical innovation team. Where we're going, we don't need roads

[futurekeys-blog]

We bet you always found yourself in a situation where you couldn’t figure out from where a person was, based on his/her accent. So, we challenged ourselves to that. To detect from where is a person based on the accent in a common language, english.

This will not only solve the german intruder from Tarantino’s movies, it also can help speech recognition engines and voice assistants.

[futurekeys-blog]

Lean Data Scraping

We often need to retrieve data from the internet for the purpose of creating a dataset. The source can be a REST API, webs, other databases or the hated XMLs. So that can be painful. We try to be lean doing that, so we use some fast prototyping tools for that. Python, along with modules like requests for APIs, scrapy, BeautifulSoup or directly tools like import.io for web scraping. If the size of the JSON/CSV gets larger than a few KBs, you may think of getting some database involved, like MongoDB. In that way you can avoid losing your data and performing stress test on third party APIs. 

“But using MongoDB is not lean because you have to install, start and maintain a mongo server... bla bla bla”. Don’t worry, you can use a Database as a service like mLab (ex MongoLab), so you can do it with a few clicks. And also allows you editing the data if you are not comfortable with the console.

[futurekeys-blog]

4YFN attendees

Map created by in CartoDB

Position and categorization of attendees at a big startup event.

[futurekeys-blog]

We wanted to create a map with all the technology and their hype, but our biggest issue was choosing an awesome visualization for that. The metaphor we decide to use it’s a galaxy. Were a solar system it’s a technological category.

We are using WebGL to paint the galaxy along with the Three.js library, so it could run in any browser. Also to store the data and organize it, we use Trello and the JS API

[futurekeys-blog]

How web could revolutionize mobile, but it didn’t

If you ever developed a native app and you ever used a WebView or iOS’s UIWebView you may have some bad memories. Even a simple hello world web would make your app slow, like it was running in a Nokia 3310. Not just that, things like the 300ms delay on click or iOS and the limitation of using their Web Engine. Those things made mobile web apps slow and unusable.

In that way all the power and flexibility of web it’s gone. Frameworks like p5.js, three.js, d3.js are useless because of poor performance. That’s why all the web cross platform frameworks failed miserably. The thing is that the mobile browsers didn’t behave like that. You can make an app with three.js + vr in cardboard, data, leap motion and voice recognition, working fine in a web browser.

Right now, things are changing, the webview from the Twitter app uses Chrome, the WebView componente it’s updatable from the Play Store. In Apple we have a new webkit componente, WKWebView. that gets the same performance as Safari with iOS 8 and higher. 

We should build better web and allow coding Flappy Bird just once.

[futurekeys-blog]

A lot of mobile in 4YFN, as if it was part of World Mobile Congress ;). Image from http://beeva-labs.github.io/radical-4yfn/

[futurekeys-blog]

Categories, relations and details

We made a dataviz using the startups from #4yfn. We included Accelerators making a total of 426 entities. We show categorization by industry sector, details, links, pitches. The companies are connected to each others building a graph between the keywords obtained with tf-idf.

[futurekeys-blog]

Clustering startups by pitch

Recently we tried to cluster automatically startups using their pitch. We are talking about description with an average of 52 words per description. And a total of 426 companies.

We used tf-idf (more about here) to get the relevant keywords from pitches. Next we tried two clustering methods. First k-means playing with the number of clusters. No logical results, everything mixed up, IoT, Fintech, Mobile, Accelerators, etc. Our second approach consisted in using DBSCAN. This way the clusters are auto organized by density. We got better results, but still, 50% in the same cluster and 40% of them got lost in the outliers zone.

So, clustering startups with short pitches may be a bad idea.