S tone Soup Tech is an Artificial Intelligence oriented software and consultancy agency focusing on providing strategic research and development. Our AI department is headed by:
lecturer at the Polytechnic University of Bucharest teaching Machine Learning applications
member of the Romanian Association for AI (ARIA)
member of the AIMAS-WO promotion and development team
former research developer at the Serious Games Institute (Coventry, UK)
former member of the Laboratoire d’Informatique de Paris 6
Game AI, Machine Learning, Expert Systems, Planning, Knowledge Representation and Reasoning
ranked one of the top 100 data scientist on Kaggle.com
worked on terabyte-scale datasets at Twitter
TA at the University of Bucharest, teaching Neural Networks and Algorithms and Data Structures
Predictive Analysis, Natural Language Processing, Machine Learning
designed & implemented multiple AI centric applications including statistical classifier for signal auditing lead generation, various recommendation systems, AI powered social bots
former member of the AI-MAS WO Team, implementing a platform for Game AI competitions
Natural Language Processing, Statistical Classifiers, Behaviour Analysis, Machine Learning
Health care article
It is a system that combines Business Intelligence with Health education. It improves the health-care environment through smart online education and actionable insights offered to the specific entities by making use of Artificial Intelligence and the power of the Cloud.
Our system integrates artificial intelligence (AI) technology and internet-based health information to deliver a system that indexes, analyses, and feeds health-related articles to users based on individual reading preferences.
Our first generation system has 75 indexed websites with approximately one million articles within our system on various health related topics. We can analyze data on a particular disease, a drug, or a symptom and provide data summaries. These can be used by pharmaceutical or health-care consumer companies in the pre-discovery phase of drug/product development when understanding all facets of a disease or disorder are critical before embarking on further drug and or product development. Both pharma and consumer companies can capitalize on a detailed understanding of consumer trends. Analyzing our user base for patterns and trends in most frequently read articles, most shared and most commented on articles can provide valuable information on future market trends. This information is otherwise difficult to capture and our system provides a unique opportunity for pharma and consumer-based companies to gain a better understanding of their costumer.
Real estate recommendation
& analysis system
This system uses Artificial Intelligence to provide fast insights and drive agile, competitive decisions and actions for the real estate businesses, while helping regular users find what they need faster by using analysing to their behaviour and preferences.
Our solution brings Artificial Intelligence into play in the online and mobile real estate market. It harvests data from the web as well as user behaviour from mobile apps. Combining these two data sources with data mining solutions and machine learning techniques, we can provide valuable insights and analytics, as well as uncapturable information for the real estate industry. Our intelligent system recommends the right listings to users within minutes of utilizing our applications.
We make use of knowledge fusion and statistic techniques in order to match social media profiles in different networks to the same user (for example figure out if a facebook and linkedin profile belong to the same user by overlapping the social graphs of each account).
Based on the input received from a user the system searches for profiles and calculates a match probability score, providing the most probable results. This system was requested by government agencies and collection agencies to generate case leads.
We’ve researched a machine learning approach to predict the success or failure of a campaign and found that it has a performance of 67%, somewhat higher than the Kickstarter average of 54%. As expected given the particularities of the data (short feature vectors, large dataset), the most successful standalone algorithm is the random forest classifier. We prototyped an ensemble using an RFC, along with logistic regression on the same feature vectors (these algorithms complement each other in avoiding overfitting) and support vector machines on the text data (SVMs are very good on sparse data such as text) would improve the result.
If we factor in features related to the first few hours after the campaign starts, an algorithm can predict the success with over 80% confidence. We can notice the results getting better as these features get richer. We’ve also found that analyzing the virality factors behind social interactions and user backing are key in providing accurate and actionable predictions.
This system was one of our first AI based applications. Our client required a system to identify pubs and sports bars in the United States that advertised specific sport events and shows on the web, without owning the appropriate broadcasting license.
For this we built an indexing engine querying all the major search engines as well as the major social & blog networks used in the US. All the returned pages were indexed and analysed using Natural Language Processing and image matching. A match probability score was computed and based on it leads were generated. Following that the site was indexed, the contact page identified and using NLP and regular expressions one or more addresses were extracted. These addresses were compared against a client list provided by the client.
The address list provided by the client contained over 50,000 addresses and client entries. Searching and finding leads manually would have taken a person thousands of hours while our system provided new leads day over day.