ProblemRecostream came with a mission to make advanced recommendation models available to e-commerce sites and content platforms of any size. Development of the new system has to be strictly data-driven and rigorously tested in terms of recommendation accuracy and latency. Integration with any eCommerce platform, a content system or a mobile application should ideally not require any technical resources on the client side. The performance of the recommendation engine should also be open for validation with third-party tools.SolutionWe have researched and improved a number of AI/ML driven recommendation models. The designed  integration method provides quick integration for non-technical users by installation of a single Javascript import line that works with any eCommerce platform. More complex integrations are available with Advanced API. Powerful dashboard capabilities were extended with complete integrations with Google Analytics and Google Optimize what enables clients to verify the actual, objective impact on your store’s metrics within GA and run comparative A/B testing with no development resources needed.ResultsRecostream was acquired by GetResponse, a global online marketing leader, in December 2022.The actual, real metrics show that most eCommerce stores should observe actual sales boost by 5-10% with Recostream recommendations.Deep integration with Google Analytics enables store owners to track recommendations performance within the context of other metrics with confidence and trust.The system despite super high event volumes returns a new recommendation within 20-30 ms. Learn more about how recommendation system works.

ProblemThe objective was to add more context to the existing data stored in a Salesforce instance by enhancing it with additional information queried from multiple external systems. The idea was to collect the maximum scope of information within Salesforce for further machine learning analysis without any major prefiltering. The APIs of external systems were mostly documented in RAML.SolutionTo enable the business in their exploration of contextual data from third party systems that could improve the prediction power of their custom machine learning model we decided to build an automated load tool that would enable the client to easily add new data feeds. Our aim was to simplify the whole data model creation and data load process and treat Salesforce Platform as a general purpose data store. Putting MuleSoft integration platform into action we created a universal mechanism that automatically generates a custom Salesforce data model on the basis of RAML specification of the API and then populates it in a full snapshot or incremental fashion.ResultOur proprietary API connectivity tool enabled us to turn Salesforce into a general purpose data store and to quickly populate it with contextual data from various third party data sources. The automated process of Salesforce schema creation and data loading dramatically improved the business ability to try and explore new data sources for their predictive machine learning models.

ProblemA global flight information company run into scaling issues with their existing schedule calculation system. The existing solution was a mix of DB2 in-database processing and a proprietary Java code. Performance scalability was limited by both system design and commercial license costs.SolutionStratoflow proposed a PoC that would replace the implementation of the major calculation logic with an open-source, high performance framework. A short series of spike iterations was planned to design and implement the selected logic and to run comparative performance testing using exactly the same input data as the existing production system. A small team was created with Stratoflow consultants and a development lead from the client.ResultThe prototype design and implementation was confirmed by running functional tests that confirmed the daily processing output between the existing production system and the prepared POC solution. That enabled us to confirm that we will be running apple to apple comparison looking at performance results of both systems.Comparative performance tests were run between the existing production system and the POC solution that end up with three major conclusions:POC code was faster than the existing production code.Open-source technology selected for POC implementation had dramatically lower TCO when compared with the existing code.The new design is easier to developer, maintain and deploy than the existing system based on obsolete technology mix.