Сестра наша внатуре.

ЖЗЛ!!! ЖЗЛ!!!

что-то мне подсказывает, что это не согласовано с УГПС ГУВД (МО), Мос(обл)энергонадзором, Мос(обл)комприродой, Комитетом по ЖКХ (МО), Мос(обл)инжстроем и МосгипроНИсельстроем.
так нельзя.

Приходите прятаться к нам под ванну.

Kuda kota ?

Сергей, я копирую письмо сюда, потому что Ваш адрес ответил мне отлупом. Вы это удалите, пожалуйста.


Сергей,

здравствуйте,

мне прислал ссылку на Ваше резюме мой старый знакомый Саша Николаев (http://www.livejournal.com/users/ati/). Я ищу полно- или частичнозанятого исполнителя вот для такой и подобных работ (текст ниже). Нет ли у Вас знакомых, включая Вас самого, которые бы для этой работы подошли, и которая бы их заинтересовала?

Давид Толпин
http://davidashen.net/

Our company collects names, addresses, and other information from its respondents through a Web survey. We attempt to match these respondents to their voter registration records. Almost every state in the U.S. requires voters to register to vote prior to an election. We have built a database containing approximately 160 million voter registration records. We use the voter registration data for two purposes: (1) to have data on the voter's party registration and past voting history; and (2) to make it more difficult for a person to take surveys multiple times (by checking that there actually exists a person with the name and address supplied by the Web survey respondent). We try to process the respondent's name and address information in real time (as they are taking a Web survey).

 
Each record in the voter registration database contains the name of the voter, their address (including city, state, and zip code), and usually their birth date and gender. States which are covered by the Voting Rights Act also have the voter's race. Other data in the voter registration records include turnout information (whether a person voted or not; who they voted for is, of course, secret). In about half the states, voters declare a party when they register to vote and this information is in the voter registration record, too. In states with "open" primaries, we usually have information about which party's primary the person voted in. All of this information is very useful to us in constructing representative samples of voters.
 
We use a multiple step matching process:
 
(1) First, before we have asked the voter for his or her address, we ask for their name, zip code, gender, and birthyear. We then look for a matching voter in the voter registration database with the same last name, zip code, gender, and birthyear. If there is a single match, we verify that the first name on the voter record is approximately the same as the first name provided by the survey taker. If there is more than one match, we check to see if there is an exact match in the voter list for the first name supplied and name suffix (e.g., Sr., Jr., III, etc.). (U.S. zip codes contain approximately 10,000 persons. The procedures above work reasonably well for distinguishing between fathers and sons and husbands and wives. They do not work well for common last names and are sensitive to small variations in spelling and nicknames.)
 
(2) If the first step above did not produce a satisfactory match, we then ask the respondent for their address. Because there are many different ways to write an address, we "standardize" the address and check to see that such an address is listed in the USPS Zip+4 and DPV files. We have developed our own Python library for address standardization. We maintain the Zip+4 and DPV data in a PostgreSQL database. (The USPS databases are about 1GB total in size.)
 
(3) We then ... (места не хватило, лимит 4300 символов)
 
Tasks for this project:
 
(1) There are many limitations to FEBRL, including a poorly designeed architecture. We would like a strong Python programmer to assess what is the best strategy for using FEBRL (to write some functions to wrap it, to call some of the functions, or to develop a simpler library borrowing ideas and code from FEBRL). FEBRL supports some fairly complex algorithms for processing matches using hidden Markov models.
 
(2) To calibrate the parameters of the matching process to improve match rates, while controlling error rates.
 
(2) To improve the performance so that matching can be done for a large number of simultaneous users in real time.
 
(3) Using XML-RPC or some suitable technology, to create a service that can be called from our panel registration system.