• AIforSEO
  • PlayLy
  • GrainMonitor


An app for evaluating and improving content quality for better SERP performance

Problem / Purpose Statement

To analyze web page textual content and subsequently rank search results, Google uses a machine learning model based on a neural network architecture called Transformer.

This Transformer-based ML model analyzes the text as a whole, without sticking to the exact keywords: how relevant it is to the user's intent, how well the topic matches the query, etc. — basically everything that can help to provide the most accurate answer that meets the user's needs.


We tried to adapt to this model by building the same one but hundreds of times smaller. Utilizing the transformer architecture, we trained our model to automatically adapt and duplicate the original algorithm. So that is how AIforSEO was born. Now, our system can show you how Google will classify your web page content and how high it will rank it on the search engine results page. Everything is processed on virtual servers.

How Did We Do It?

With the help of Google-like AI and neural network-based models, the app analyzes Google's top-10 results relevant to the search and your website's text content. Then, it compares the collected data and, based on this analysis, shows what is missing and what changes you need to make to your content to improve its relevance to the query — the main goal of modern SEO.


AIforSEO analyzes and compares texts so that you can use the resulting report to edit and improve your content without affecting its readability.

The project received an honourable mention in the Artificial Intelligence category at the Go Global Awards 2021.

Target Audience
  • English language-oriented companies with high-frequency queries
  • SEO agencies with English-speaking/English-oriented clients
  • Marketing & PR specialists, freelancers
  • Small businesses that want to do a competitive analysis

A prototype demo of AIforSEO has already been released. Free slots are provided for those working with the English language and Google (email us at seo@aiforseotools.com ).



A match tracking-based football analytics research project

Problem / Purpose Statement

There is a common perceptual problem shared by both fans and experts: it is not clear how to evaluate individual player contributions in such a team sport as football. The team as a cohesive whole is 11 players organized into a certain structure. We can tell when the team wins, draws or loses; but can we measure an individual's impact on the match's outcome? And if so, how?

For example, a transfer fee for a player is determined at a rough guess, simply through an agreement — there is no special rule, algorithm, or formula. The issue doesn't stay completely unaddressed as new analytical tools are emerging. But in general, the problem of correctly evaluating football players’ performance has not yet been solved. So, we set a goal to work towards the development of a solution.


To build analytics based on player performance data, one needs to collect these data from football broadcasts. Videos from YouTube broadcasts are not enough since they offer poor field view, the image scale is constantly changing and not everything fits in the camera frame. Recordings from a tactical overhead camera give a more accurate picture, but these are not available in the public domain.

How Did We Do It?

We gather data on a particular match from several different broadcasts: standard camera footage (different angles and quality) and tactical camera footage if we can find it. With the help of OpenCV computer vision algorithms and the PyTorch neural network model library, we transfer the data of the chosen match to the field — basically converting 3D data into 2D data and reconstructing the exact locations of each player and the ball during the match

One model recognizes the football field in its entirety or by key points, another model recognizes the ball (which can be less visible in low-quality videos), the third one recognizes the players — the results from the three models are then combined to recreate the pipeline. The data on players' movements undergo mathematical analysis.


Advanced scoring algorithms that assign some significance or value to each movement of a player are too closely tied to what this player does with the ball. The degree of opposition from the opponents and the support of teammates who do not control the ball at the moment are not taken into account. However, other players' positions in the field, their movements, and actions are also significant.

On average, a player has a total of between 120 and 180 seconds of ball time in the entire 90 minutes of play. That means each player spends most of the game working off the ball. And they do work even if they do not control the ball; it is not just hanging around the field — they act, move, and communicate with fellow players. These off-the-ball actions affect the course of the match and its outcome but are very difficult to assess. So far, there are no tools for this kind of assessment — only a few scientists in the world analyze off-the-ball models. That is why our main focus is evaluating the actions of players that do not possess the ball.

Target Audience
  • Football coaches and players that want to improve their performance, evaluate their performance in a particular match or even some part of a match, find the best tactic to play against their opponents.
  • Football scouts that want to adequately assess a player's transfer value
  • Football clubs that need to form a team for a concrete task



A document management system for agricultural holdings / Crop price forecasting system based on data monitoring

Problem / Purpose Statement

It is necessary to simplify the document management and communication on both sides for large-scale transactions when agricultural products are purchased simultaneously from different suppliers. At the same time, the volatility of prices for agricultural products hinders the formulation of the company's short-term and medium-term financial strategy. With quick decision-making, it is better to rely on analytical reports covering market conditions, crop prices and other influencing factors.


We organized data collection from various sources (news, social media, stock indices, satellite images showing fields and ports) by type, organized report generation, and built mathematical models for predicting urgent demand in the intervals of 1–3, 1–5, and 1–7 days.

How Did We Do It?


  • Fact parsing & extraction module
  • Demand estimation mathematic module
  • Satellite data and visual information analysis module

APIs and systems:

  • Market conditions monitoring systems
  • Exchange systems
  • GIS systems
  • Weather systems

The project is protected by NDE.