Adam Seever

Welcome to Adam Seever's web projects page. The following are some of the more interesting things Adam has created over the years. For a closer look into Adam's thoughts on topics various and sundry, check out his posterous site.

Currently in development, Make and Store is an inventory forecasting system for small businesses. Inspired by some of Adam's client work, Make and Store seeks to provide a lightweight SaaS framework for anticipating shortfalls and recommending supply chain decisions for small businesses who want to track in-process manufacturing states. Drawing from Adam's experience developing history-tracking solutions for transactional systems, Make and Store's most interesting architectural feature is the ability to track historical state changes on mapping tables yet still operate within Rails' ActiveRecord schema. Though not ready to take customers the daily "dog food" release is online and available for browsing.

Now in its third major version, Semantic Gifts is Adam's most significant solo project to date. A text mining engine which uses Twitter and other social media sources, Semantic Gifts seeks to classify users based on their interests, how they view gifts, and the way they use Twitter. Recommendations are driven by a machine learning algorithm that uses tiered support vector machines arranged in a "hidden layer promotion" configuration influenced by neural net architecture. The filtered TFIDF vectors on the NLP side demonstrate Adam's extensive experience cleaning up dirty social media feeds. The initial version released for Christmas 2008 used Latent Dirichlet Allocation, Wordnet and Link Grammar to do topic modeling, sentiment analysis, and named entity extraction. The next version released for Valentines Day 2009 marked the introduction of a cross-referencing system for matching gifts to personalities (over named interests) and personalities to Twitter users as well as a comprehensive back-propagation system. The current version uses layered support vector machines to cross reference interests, twitter style, and perspective to make the best recommendations yet.

Towncry was built as a dashboard for finding interesting things to do in Seattle. It was the first project Adam worked on after arriving in Washington, and was a good way to settle in. Though officially based in Adobe Flex, the majority of the code in Towncry's final version was written in custom ActionScript. Account management was handled by .NET.

Adam's first ML/NLP project, Morning Set was an OPML based news recommendation service. Users could upload their RSS news feed lists, weight and organize them, then receive filtered streams through new category feeds. The site was written in Ruby on Rails and took advantage of OpenCalais for determining stories' keywords. The user comparison feature used a simple bayesian algorithm to determine distances between users and the reader used custom parsers and stop lists to scrub the OPML docs and RSS feeds.

A fun side project using the Semantic Gifts engine, this app sought to identify mentions of drink related words in order to determine what beverage to order or bring to the home of the Twitter user in question. Though some menu-scraping experiements used other technology, the prototype was all Ruby on Rails.

A side project to apply some of the ideas Adam formulated for getting around on foot, Adam wrote up and coded an algorithm which beats major map sites' walking directions by 10% or more on an average trip. The initial version used Python, Django, and Google App Engine, but the final version was written in Ruby on Rails.