I am giving a talk at the local Node user group on building a life embetterment robot. This is my current outline.

how_to_train_your_robot.md

How To Train Your Robot

Not that kind of robot

This kind of robot

What is this robot you speak of?

Enter: Hubot

What is Hubot?

• Hew-bot
• A Customizable, Kegorator-Powered Life Embetterment Robot
• Written in Node.js and Coffeescript
• Uses Redis to persit data in its "Robot Brain"

What is Coffeescript?

• The Devil
• CoffeeScript is an attempt to expose the good parts of JavaScript in a simple way.
• Written By Jeremy Ashkenas
• Compiles to vanilla Javascript, with no interpretation at runtime

So what do you use Hubot for?

• Originally as a chat bot and to automate running tests etc.
• Generating memes
• General time-wasting

How do you use Hubot?

Adapters

• connects Hubot to the outside world

Syntax sugar that is coffeescript

Scripts

• Repo managed by github
• stuff the robot actually does

Script

• Two ways .respond and .hear, respond requires a prompt (primary way)
• send the message out

But wait there's more!

• I started thinking about how limiting pre-built life-embetterment robots were
• Hubot is command driven, Siri and Google I have little control over their tasks
• What if I could merge the two?

Enter: Natural Language Processing

• I started looking into how I could get Hubot to understand me more naturally
• My wife's a Master of English so I began asking her to re-teach me english syntax
• How do I get the computer to understand English?
• Tokenizing, trying to find relevant words etc.
• Started building my own language processor...really, really hard.

To the Googles! Da na na na na na

• Started looking for open source natural language processors for node.js
• Not nearly as roboust as I wanted

What I found

• Stanford Language Processor
  • Java App
  • Had node run it in a sub shell and pass it stuff
  • Really slow
  • Returns garbage, so I needed to create an adapter to conver to JSON

Still not what I wanted

Enter: Elastic Search

• Combined with the HTML Web Speech API
• I could now just send the entire string and get weighted results

So How'd you do it

• Well naturally I named my robot Jarvis
• And naturally I wanted him to speak in a British accent

Enter: ElasticSearch

• I could now just send plain english to my database and get weighted results
• results will have a string that will match one of the hubot scripts regex, low manipulation of hubot core.

Enter: i18next

• Run all of the robot's responses through tranlsation
• Sectioning out the parts of speech
• Localizing responses to time of day and randomizing similar response phrases

What Does It Look Like?

• SQLite
• Elastic Search
• Rails
• Node (Hubot)
• XMPP

Why Rails

• I looked at node options for elastic search apis and found that for speed of development and general niceness the Tire gem was great.
• Good Practice

The Gift of Speech

• Hijack Apple's speech files to get a cool British accent
• Have node execute say commands in a sub shell to get the robot talking

Spike: Language is inherently synchronous, Node is not

• The robot was talking over himself
• Quick touch on the async library and managing control flows with node
• Show stupid binding in coffeescript

Nitty Gritty

• XMPP adapter for Hubot for use with a web interface
• Jarvis talks to his "Brain" which is a SQLite with a rails server on top
• uses XMPP
• Why Rails? Because the Tire gem is amazeballs
• The "brain" responds with an "executable" object which node will then parse the json and execute the function contained in the object.
• Still uses is lesser Redis brain for simple things

What's next?

• I would like to build a more robust web interface for him so that he can respond with visuals as well as audio and text

Quick Demo

Yeah scripts would be good.

Forseeable questions, putting on the user hat:

1. What's a Redis? Is it actually red?
2. So what exactly is i18next? Something to do with translation, you say?
3. Diagram outlining the workflow of the bot ('Nitty Gritty' in a picture rather than just bullet points)?

Also, for presenting, I really dig the Reveal.js that @mrDarcyMurphy used for his presentation last week http://lab.hakim.se/reveal-js/#/

Gonna be schweeeeeet.

I can't believe your hubot's name isn't german.