Clustering Algorithms Developer in Sofia, Bulgaria

CHIN hold a large database of artifacts coming from museums all over Canada. This database was obtained by collecting all digital datasets that the museums provided, which are end-of-life, meaning that they are not maintained anymore. In its current state, the data collection is hard to use, suffering from ambiguities (eg. many spellings of same author name), repetitions, mixed and unresolvable language input (EN or FR) by automatic means, lack of standard object taxonomies (eg for object materials, types). Ontotext was given the task to evaluate the effort necessary to clean the database and link to LOD resources (DBpedia, Getty AAT, and others).

I worked with two colleagues on this project. My role was to statistically estimate the proportion of malformed data, focusing on its most important features (eg. museum, objects category, type, name, language). I also had to estimate what proportion of the errors are systematic and are addressable by automatic methods (NLP).

Eventually, the project was successful, exceeding the expectations of the Canadian institution.

In the weeks preceding the Brexit referendum, Ontotext streamed discussions from Twitter on the topic and ingested them in GraphDB, using the PHEME (www.pheme.eu) ontology model and the semantic enrichment pipeline, which affords the linking of entities to LOD data like Geonames and DBpedia. I was assigned the task to analyze these tweets and determine which are the main actors of the British referendum and their stand w.r.t. the main question: #leave or #stay? By using some simple bootstrapping techniques around polarizing hashtags (eg. #leave, #exit; #stay, #remain), I was able to conduct sentiment analysis on the set of tweets, which provided with an estimate of sentiment for various types of entities: politicians, political parties, geographical locations., age groups. My analysis showed that the lobby for #exit was much stronger than for #stay, at least on Twitter. The report was published days before the referendum ( https://ontotext.com/twitter-users-support-brexit/ ) and it received a lot of attention after the voting result matched the outcome of the analysis.

I had the curiosity and ambition to find out how easy it is to implement an algorithmic high-frequency trader for the cryptocurrency market. I conducted this project by myself. The technology employed was Python and Keras, and the methodology was Reinforcement learning with deep neural networks. I worked with open data, namely three-years prices of Bitcoin. I also generated synthetic data, for testing of the algorithms. I have educated myself on the topic of reinforcement learning and time series forecasting.

PHEME was a very successful research project funded by the EU’s 7th WP for research, technological development and demonstration. The consortium had the ambitious goal of developing a platform for automated detection of rumors on social media, intended to help journalists fight the proliferation of misinformation and disinformation through such channels. The tool was streaming data from Twitter, filtered to cover interesting political events. The data analysis was a multi-language text-processing pipeline on top of an ontology that modeled rumors on Twitter.

I was involved in many aspects of the PHEME project. As a data scientist, I developed an ML model for prediction of rumors on Twitter. As a member of Ontotext's team, I coordinated the integration of various pipeline components coming from all partners. I wrote deliverables, reports and scientific papers describing our work.

I have been hired by MobiBiz as a freelancing data scientist (through Toptal) to implement and bring to production a system that is able to extract relations from highly structured documents. These documents include tables, sections, figures with captions, etc. The solution is based on the open-source Fonduer algorithm. My task is to apply this algorithm to a specific dataset but to keep the solution general enough to ensure applicability to future similar cases. The code is in Python and the application uses a Postgres relational database, with the SQLalchemy interface to Python.

I helped with the implementation of two chatbots, depicting characters Alice and Cheshire Cat from the book The Adventures of Alice in Wonderland by Lewis Carroll. The characters can interact with users who ask questions verbally. A speech recognition system translates voice to text and a question answering system is providing an appropriate answer. For this prototype, the character has a fixed set of responses to choose from, referring to facts from the book, biographical information on Lewis Carroll and his books, and some topics about the University of Southern California’s library.

My role in the project was to help my team select a speech recognition system that can be used to translate users’ questions into text and to implement a question answering model that is able to select the appropriate answer from the list of possible answers. I used BERT for question answering. The system is deployed as a web service and takes requests in real-time, through a Flask app.