Peter Doan Email & Phone Number

Alameda, Ca, Us

Alameda, Ca, Us

San Francisco Bay Area

Data.com, salesforce

Fredericton, Nb

Main duties: delivering high-value, low-latency, high volume and high quality data at scale to our customers through thoughtful work and innovationMajor projects:- High Performance Keyword Matching: Maintain the Zapp boolean keyword matching system running on GPU, massively parallel processor platform, using CUDA and Python, matching 6-7K documents per second using a single core.- Search Relevance Ranking: Completed research and proof of concept to improve search relevance ranking using Sphinx. Improved Precision@n score up to 0.7 using customized rankers in Sphinx. - Japanese Spam Classification: Built a super fast spam classifier for Japanese webpages with 0.3% false positive and 4% false negative. Further combined training model with ad-hoc methods to reduce false negative rate to 1.6%, caught about 8000 spam webpages per day; each takes 80 milliseconds - Spam Blog Detection: Developed multi-layer spam detector framework from latest research work results on spam filtering, such as spam regex patterns, linguistic analysis, content compression rate, link analysis. Further invented new algorithms to prevent spam flood and applied for patent consideration. Reduced overall level of spam in blog data in system by 53%, caught about 302,800 spam messages per day.- Aggregator Detection: Aggregators are websites that steal content from other sites with no indication that their contents are not original. Designed algorithms to correctly identify 97% of aggregators with 1.8% false positive rate, caught about one million aggregators per day. Scaled system gracefully to handle more than 3 million messages per day.

Canada

Main duties: Member of Noise Reduction team to deliver industry-leading volume, breadth and depth of data, effectively classify all noise types and keep latency to a minimum.Major projects- Noise Classifier for Twitter content: Designed and implemented algorithms correctly captured 98.5% of noise with false positive rate of 1.2%, caught about 865,370 noise records per day.