James Hammerton Email & Phone Number

Hannover, Lower Saxony, De

London, Gb

• Investigating methods for developing embeddings for documents and queries that can be used for matching questions with documents that contain their answers to see if we can improve on the existing model• Helping drive development of a library of common code for implementing models, providing easy access to datasets and to the related preprocessing that the company uses. • Helping create a Wiki documenting Kare's models and the results obtained with them, as a record for future reference.• Exploring how to detect when a text classifier is presented with inputs outside of the domain of the documents it was trained on, to see if one can detect when to search a wider database vs sticking with the model's predictions.• Implementing a text classifier to decide which of a fixed set of documents a question is most likely to be answered by.• Exploring ways of clustering questions to aid detection of paraphrases, and identify potential subjects not addressed in databases of customer support articles.• Helping develop a proof of concept for extracting instructions from FAQs with a view to building processes that guide users through the instructions.• Exploring the use of word vs character embeddings, and the impact of noise levels on both cases, for the text classification model below, leading to a published paper.• Helping design and develop the text classification model that was used in the company’s old task manager application to aid in detecting and surfacing the tasks that appear in users emails.• Developing a model to predict when a user might reply to incoming emails.• Developing a Spark pipeline for extracting features and datasets from emails that could be used to predict how users might handle those emails.• Helping to design and implement features in the company’s early stage ‘personal assistant’ application, exploiting ‘related files’ functionality.

London, Gb

At VisualDNA, my work has included:* Developing metrics to evaluate the performance of VisualDNA's Hadoop-based "lookalikes" algorithm, used to target advertising campaigns to users whose profiles are similar to those of users known to be relevant to the campaign.* Analysing data using Scalding, Hive and/or R to help understand performance of lookalikes.* Exploring and developing new algorithms for lookalikes.* Integrating the lookalikes code base with the main code base for other inference algorithms to facilitate code reuse and faster development.* Planning work for a small sub-team dedicated to lookalikes.* Performing Skype and face to face interviews with potential recruits.* Advising sales team and clients on best use of lookalikes data. * Developing a prototype service to extract keywords from web pages for use as features in machine learning, using Cassandra, Kafka, AlchemyAPI (a third party API for extracting content from web pages) and multi-threading.* Developing code enabling efficient lookup of keywords from Cassandra extracted from web pages during feature extraction.* Helping redesign and reimplement lookalikes code to support performing multiple lookalikes jobs in one run, plus simplifying code to point where lookalikes and inference are basically different configurations of the same code.* Analysing some third party data to see what impact including it would have on the quality of inference algorithms.

London, Gb

At Mendeley I worked on the following, mainly Java projects:* Designing, building, maintaining and extending their deduplication system. This generates Mendeley's catalogue of research papers, supporting features such as recommendations and search. It handled over 50 million articles, generated from 150 million documents uploaded by users or imported from external sources.* Developing the initial version of the "related research" finder, using Lucene and MapReduce* Migrating the deduplication system from MySQL to HBase.* Developing MapReduce jobs for merging "missed duplicates", reclustering documents in the wrong cluster, gathering performance metrics and performing metadata aggregation.* Creating tools to generate deduplication datasets for evaluating new algorithms with and the means to run the deduplication system in a testing environment.Additionally, I helped mentor new members of the data mining team and gave talks on issues such as test driven development, the use of mocking and how to unit test MapReduce jobs. I contributed team planning and participated in interviewing prospective team members.From mid-July to mid-October 2011, I was seconded to the Know Center at the Technical University of Graz, Austria for the TEAM project. Here I expanded on the evaluation tools and generated large datasets for deduplication, and established that a Solr based approach can work as well for deduplication as the existing service.

Helped design and implement ciboodle's agent247(TM) technology, allowing natural language dialogue with computer applications via speech (using VoiceXML), text chat, email and SMS input. Helped produce and to give demonstrations of the product to customers.Investigated feasibility of automating contact centre management tasks via machine learning/AI.Helped integrate product with Nortel telephony servers.

Investigating the use of self organising maps for cognitive modelling of the learning of speech segmentation, plus delivering tutorials in machine learning and inferential statistics.

Applying neural networks to the shallow parsing of natural language as part of the Learning Computational Grammars project.Co-edited the Journal of Machine Learning Research, Special Issue on Shallow Parsing (published 2002).

Phd, Computer Science

Bsc(Hons), Artificial Intelligence And Computer Science