Spaceport

We have developed a backbone of the Machine Learning platform for financial quantitative analytics. The core problem of existing manual analytics modelling methods is that they take a long time to build a model, while the data remains relevant only for 30 days or even less. Thus said, building a model should happen fast in order for it to remain relevant for long. This is a huge market with tremendous daily transactions activity.

We developed a backend part of the platform using neural network that stores lots of the model components for financial quantitative analytics. The user is able to quickly compose a model from blocks using a GUI, and the backend creates the specified model within the platform.

Using our model constructor helps create, adjust and manage quantitative analytics models on the fly. Our platform helps people without any ML knowledge tap into the ML power. This brings in the competitive edge and helps make impactful decisions faster, which is especially important in the financial analytics world.

There is another solution we built from scratch for a high-tech start-up — an innovative job board for Asian market. This job board must serve the healthcare and medical science industry and is oriented at top-tier researchers and scientists. There was a dual challenge — both the employers and the applicants cannot describe their skills precisely (either because they cannot lift the NDA or the project descriptions involve highly specific terminology that does not have a commonly accepted meaning yet). Therefore, the job board should be able to understand if the applicant’s skills are relevant to the job posting requirements and also recommend the applicants based on additional parameters. even if there is no exact match of their skills with the job requirements.

We have developed a fully-functional back-end  prototype and are currently working on a GUI. The job board works the same as some other popular job portals — the applicant remains anonymous until they privately provide their contacts for the interview, so their current employers have no way of knowing their personnel is actively seeking for a better position. Thus said, our prototype addresses two main concerns of all job boards:

• Privacy
• Skill-relevance

The prototype deploys Word2vec system — a natural language processing algorithm that analyzes both the job postings and the applicants’ CV’s with the help of PoS (Part of Speech) tagging. This helps find correlations between the job requirements and the applicant’s skill summary. The process is augmented with our own algorithms and tuning. Reinforced learning of the ML model happens at the moment of Word2vec suggestion, so each successful skill match or addition of a new skill to the base improves the accuracy of suggestions.

This solution provides the following benefits:

• drastically reduces the time and money expenses on applicant search
• augments the applicant screening to highlight the most relevant applicants
• shortens the recruiter’s overhead
• allows augmenting the recruiter’s soft skills with the platform’s features.

One of the main problems the analytics department in corporate business faces is the sheer volume of incoming and outgoing data, meaning the analysts have to write LOTS of text (most of which is never read). This is particularly painful for writing monthly or quarterly reports, especially if they are intended to be actually useful. We came up with a solution that lowered the quantity of human errors by 58% and provided a 71% faster report writing.

One of our customers needed a NLG solution to simplify the process of report creation. The main obstacle was that corporate business has to process a myriad of reports from underlying departments that do not have unified form and structure. We designed a PoC (proof of concept) – a tool that transforms MS Excel spreadsheets into text.

The spreadsheets can be of different length or width and can contain the data in varying cells, yet the algorithm must recognize and mark all the parameters.

It begins with data normalization and processing in order to allocate all the important data in incoming spreadsheets. Then the data is uploaded into the Template generator that works by Yes/No/Equal principle:

1. if the new value is bigger than the previous, use the template “xxx increased”
2. if the new value is lower use the template “xxx decreased”, etc.
3. if the new value is essentially the same use the template “xxx is equal”, etc.

Creating such simple templates and building the trees of them allows to describe a huge variety of situations with a handful of simple parts that can be combined in any construction.

This helps create the  custom-tailored multilayer report automatically with all the details already filled in. This simple tool helps deliver diverse and informative reports in minutes and ensures 58% decrease in human-errors with 71% faster report preparation.