Highlights

• Collective Intelligence Programming
• Worked with Bayesian Filters
• Built Gmail integration.
• Worked with message queues and asynchronous processing
• Introduced Scrum process for development management

Gist

Gist started out as an experimental project at Vulcan Labs, headed ultimately by Paul Allen. When I began working on the project it was very fuzzy about what exactly the project would be. The idea of Gist was to glean important and relevant information by importing and analyzing emails. Initially, we analyzed who sent you emails and who you sent emails to, providing us with a way to generate a very rough social graph based on your email usage. From there we tried to find relevant news and blog items based on your contacts, so that you could be more informed about the people who you meet. Think google searching someone you know only having it be automatic and for all of the people in you inbox. When I first started working on Gist it was a very fuzzy project in terms of what exactly Gist would be. We spent a lot of time brainstorming and prototyping out ideas. I tried applying collective intelligence algorithms to see if we could find better information using a rough semblance of AI. The final conclusion I reached was that although the collective intelligence algorithms could produce some impressive results, I couldn't get them to produce 100% accuracy, and without that level of precision it was difficult to see how the application of them would be valuable. Close didn't cut it in this case. Still, we were able to generate enough of an idea that the initial skunk works project was rolled out into a separate startup company that continues today developing Gist as a startup, the project as of this writing is in a closed beta.

Works

• City Winery
• Gist
• Divvy
• Ringjacker
• USA Wireless
• Collectview

Programming Collective Intelligence

undefined tag `assets'

Amazon.com to their credit has done some really amazing things with the types of algorithms described in the book "Programming Collective Intelligence". The first thing to consider with these types of algorithms is that most of them depend on the network effect in order to be accurate. The more people you have the better the results will be, so this can be tricky for new sites that don't have a lot of traffic. Second is that the results are always different degrees of fuzzy but never are they 100% accurate. In the case of book recommendations you can be inaccurate but still helpful in other circumstances being close is still just another form of being wrong. Finally each algorithm does something very different than the rest, picking what algorithms to use should rest solely on what is most helpful to your application's value proposition, rather than the sense that something is cool.