This paper introduces an AI-driven project management framework designed to enhance the evaluation of individual performance within collaborative environments. The core of the system lies in its ability to continuously analyze user activity data and transform behavioral patterns such as task completion rate, time deviations, communication activity, and task complexity into interpretable and adaptive performance scores. The framework integrates an intelligent analytics module based on machine learning. The resulting model allows for real-time scoring and explainable feedback, supporting data-driven decisionmaking in dynamic project settings. By utilizing interpretable AI and modeling feature interactions explicitly, the system bridges a critical gap in modern project management tools - namely, the lack of personalized, explainable, and adaptive evaluation mechanisms. The integration of this framework enables project teams to monitor performance proactively, improve transparency, and adapt management strategies in alignment with evolving work behaviors and collaboration dynamics.
The digital transformation of project work has driven the evolution of tools for managing tasks, teams, and resources. As project complexity and distributed collaboration grow, traditional methods are being enhanced by web systems and AI. Modern IT projects demand flexibility, remote access, and intelligent automation. Integrating AI into project management enables data-driven decisions, predictive planning, and adaptive workflows – from passive tools to intelligent systems.
Recent research emphasizes the need for more intelligent and adaptive project management
tools. According to [
A wide variety of commercial platforms has been developed to support project management.
used in practice [
onboarding Up to 500 users per workspace <1 day (very intuitive) ~1 week onboarding with support 1–3 days average onboarding 2–3 weeks (complex interface) 5–7 days onboarding ~50 active boards/user; up to 10 collaborators (free)
Up to enterprise-level teams (1000+
users)
1000+ plugins, REST API,
custom workflows
fields, automation, API Extensive templates, automation builder, API
400+ integrations, advanced workflow engine
via Azure ML, no native AI
Despite their popularity, current systems often lack built-in, objective methods to evaluate
employee performance or provide automated recommendations based on predictive analytics. This
gap is recognized in recent studies: for instance, [
The proposed web-based system supports efficient project management, resource use, and productivity. It enables global access and simplifies collaboration, speeding up goal achievement. Integrated AI enhances decision-making through automated analysis and task prediction, helping assess team performance, detect bottlenecks, and refine planning.
The system is both a management tool and a strategic solution that drives results through intelligent automation. Its scientific novelty lies in applying interpretable machine learning for real-time performance evaluation, offering transparent, formula-based insights rather than blackbox predictions.
The development of a modern web-based project management platform requires a modular and scalable architecture capable of supporting various functionalities, including user interaction, task management, and intelligent analytics.
The system is based on a classic three-tier architecture (Fig. 1), ensuring modularity, scalability, and clear separation of concerns. Users interact with the system through a web interface, which sends requests and displays responses. The application server handles core business logic, user management, task operations, and AI-powered analysis. The database layer stores project data, user profiles, task statuses, and AI-generated insights. Data is transmitted via encrypted channels to ensure security, and the architecture supports both vertical and horizontal scaling based on system load.
The component diagram (Fig. 2) presents the modular structure of the system, divided into three logical layers: Frontend, Backend, and Database Access. This separation enhances flexibility, scalability, and maintainability by isolating the user interface, business logic, and data operations.
The Frontend layer includes the user interface elements such as the task board, project navigation, login forms, and real-time notifications. It interacts with the backend via an API Gateway, which handles routing, authentication, and load balancing.
The Backend layer consists of several services: user management, project and task handling, messaging and comments, AI analytics for performance evaluation and risk prediction, and a notification engine for system alerts and reminders.
The Database Access layer provides a secure interface for data persistence, managing logs and storing structured data in a relational database.
This architecture enables rapid development and allows integration with external systems, such as AI pipelines or domain-specific parsers (e.g., WDOL).
To represent the internal data model of the system, a UML class diagram was developed (Fig. 3), defining the structure of core entities and their relationships. The model includes six main classes: User, Project, Task, Board, Comment, and AI_Score, each corresponding to a key aspect of system functionality.
The User class stores participant data, including roles and login metadata. Users can be assigned multiple tasks and contribute comments. The Project class represents collaborative initiatives, with attributes such as title, description, and lifecycle dates, and is linked to multiple tasks and boards for workflow visualization.
The Task class is central, containing data on deadlines, priority, and status, and serving as a hub for user activity and feedback. The Comment class captures messages linked to tasks, facilitating communication and documentation. The Board class organizes tasks visually by stage or status. The AI_Score class enables intelligent analytics by storing AI-generated performance evaluations, connecting task outcomes to individual users for productivity insights. Each class includes attributes like UUIDs, timestamps, and enumerated statuses to ensure integrity and machinereadability. Class relationships are defined using multiplicity indicators (e.g., one-to-many, manyto-one), forming the basis for database design and API structure.
The sequence diagram in Fig. 4 illustrates the system’s dynamic behavior during a typical use case such as updating a task’s deadline, by modeling interactions between the user interface, server logic, and data storage in time sequence.
The process begins when a user modifies a task attribute via the web interface. The client sends an HTTP request to the server, which authenticates the user and validates the input. Upon successful validation, the server updates the relevant task in the database. Once the update is confirmed, a response is returned to the client. If AI-based tracking is enabled, the system may also trigger a background recalculation of the related performance score using the analytics module.
The UML diagrams collectively demonstrate the system’s modular design, internal logic, and consistent functionality. They validate its deployment feasibility, interaction flow, and data management processes.
The proposed system features an AI-based analytics module that automates user activity assessment, generates performance indicators, and forecasts task completion dynamics.
A key goal of this module is to provide objective, adaptive evaluation of individual and team productivity. Unlike manual or rule-based methods, it uses machine learning to interpret behavior in context, revealing inefficiencies such as recurring delays or resource underuse that traditional tools may overlook.
The module also detects early deviations from normal work patterns. By monitoring task flow and user activity, it anticipates risks such as missed deadlines, uneven workload, or reduced collaboration, enabling timely interventions. Its predictive function evaluates historical and live data to estimate task completion probabilities, forecast workload distribution, and identify workflow inefficiencies. These capabilities improve planning accuracy and promote balanced responsibility across team members.
The following indicators were identified for the modeling: • x1 – number of tasks completed within the reporting period (tasks/day). • x2 – average time to complete a task (minutes/task). • x3 – deviation from estimated duration per task, defined astactual - testimated. • x4 – frequency of overdue tasks (% of total). • x5 – number of task reassignments initiated or received (events/week). • x6 – number of comments posted on tasks (comments/day). • x7 – number of responses to colleagues’ messages (responses/day). • x8 – participation in non-task interactions (e.g., approvals, mentions) (events/week). • x9 – weighted task complexity index, calculated based on task priority, dependencies, and historical execution duration.
y is a continuous-valued performance score, scaled in the range from 0 to 100, where 0 corresponds to minimum observable productivity and 100 reflects the top observed efficiency under comparable task conditions. This score is intended to be interpretable by both system administrators and end users. For internal processing, the score may be decomposed into intermediate dimensions, such as technical efficiency, time management, and communicative engagement, each of which can be analyzed separately or jointly.
The dataset covers a period of 12 calendar weeks (84 days), during which the activity of 10 employees within a single project team was recorded.
The dataset includes all events related to task lifecycle management, communication exchanges, status transitions, as well as metadata concerning task complexity and deviations from planned timelines.
So, the data characteristics: m = 9, n = 7800 (total volume) nA = 2/3‧n, nB = 1/3‧n.
To construct a model for automated performance prediction, this study adopts the gradient
boosting method. The task is formulated as a supervised regression problem, where the goal is to
approximate a function f: Rn→R that maps an input feature vector x to a continuous performance
score y ∈ [0, 100]. Gradient Boosted Decision Trees (GBDT) build a strong predictor through the
iterative combination of weak learners – typically shallow regression trees – trained to minimize
the residual error of previous approximations [
Fm( x )= Fm−1( x )+ γ m hm( x )
N F0 ( x )=argmin ∑ L( yi , y^i) i=1 (1) (2) where γ m is the learning rate; hm( x ) is the regression tree; yi are the true performance scores; L( yi , y^i) is the mean squared error. Each regression tree hmpartitions the input space based on threshold splits across the features and assigns a constant value to each leaf node. These trees are weak learners by design – often with maximum depth 3–5 – which reduces overfitting and allows the ensemble to generalize well across different patterns of user behavior. Early stopping is applied based on validation loss convergence, and cross-validation is used to fine-tune key hyperparameters such as tree depth, number of estimators, and subsampling ratio. The final output of the model is a predicted productivity score, which reflects the user ’s overall effectiveness in task completion and team interaction.
The trained gradient boosting model demonstrated a strong ability to approximate performance scores based on behavioral indicators. On the validation dataset (20% holdout), the model achieved a RMSE of approximately 4.83 points, with a MAE of 3.45 points, both measured on the performance score scale from 0 to 100. This level of error is acceptable in the context of behavioral prediction, where subjective variance in human performance evaluation is inherently high. In addition to its predictive accuracy, the model yielded interpretable insights through feature importance analysis. The top contributors to the final prediction included: tasks_completed – accounting for over 35% of the total model variance, reopened_tasks and time_deviation – jointly contributing 30%, comments_posted and responses_sent – adding around 20%, with the remaining importance distributed among collaboration and complexity indicators.
Figure 5 illustrates the internal logic of one of the trees in the ensemble, revealing how the model discriminates between users with stable, timely task execution patterns and those prone to delays or rework.
X r+1=( y1r , ... , yrF , x1 , ... , xm)for a layer r+1; 2. The operators of the kind:
ylr+1=f ( yir , yrj) , l=1 , 2 , ... , C2F , i , j=1 , F , (3) ylr+1=f ( yir , x j) , l=1 , 2 , ... , Fm , i=1 , F , j=1 , m may be applied on the layer r+1 to construct linear, bilinear and quadratic partial descriptions: 3. For any description, the optimal structure is searched by combinatorial optimization; e.g.: z=f (u , v )=a0+ a1 u+ a2 v z=f (u , v )=a0+ a1 u+ a2 v + a3 uv z=f (u , v )=a0+ a1 u+ a2 v + a3 uv + a4 u2+ a5 v2 f (u , v )=a0 d1+a1 d2 u+a2 d3 v (4) (5) (6) (7) Then the best model will be described as f (u , v , dopt ), where dopt=argmin CRl , q=2p−1 , f opt (u , v )=f (u , v , dopt )
l=1,q 4. The algorithm stops when the condition CRr >CRr−1 is checked
The best variant is chosen based on the minimum criterion CR, meaning the complexity of the partial model is optimized (6).
^y =5.8 x1+0.4 x4+0.3 x5+0.1 x9 x6+0.23 x3 x7+0.2 x22 R2(nB)100% = 81%
The structure of the model includes both linear and nonlinear components, as well as interaction effects between variables, which reflects real-world processes in project management environments. x1 – number of tasks completed has the largest magnitude, confirming its role as the primary indicator of productivity. Moderate contributions are made by features related to execution discipline, such as x4 – frequency of overdue tasks and x5 – task reassignments, which aligns with the principles of team reliability and workload stability. Particular attention should be paid to the interaction terms x9x6 and x3x7 , which represent the synergy between task complexity and user engagement, as well as between schedule deviations and communication activity. These terms demonstrate that high engagement in complex or delayed tasks may offset otherwise negative performance indicators. The presence of a quadratic term x22 indicates a nonlinear effect of execution speed: both excessively fast and excessively slow task completion can reduce overall performance, while an optimal time range maximizes the score.
To assess the practical applicability of the developed predictive models, a comparative analysis was conducted between two algorithmic approaches: GBDT and the GIA-GMDH. The aim of the comparison was to evaluate not only the predictive performance, but also computational efficiency, model complexity, and interpretability – factors that are particularly relevant in decision-support systems intended for managerial use. The results of the comparison are summarized in Table 2.
The comparative results indicate that the Gradient Boosting model outperforms the GIA GMDH model in terms of prediction accuracy, achieving lower RMSE and MAE values on the validation RMSE
MAE Training time (sec) Number of model
parameters Model depth / layers
Gradient boosting dataset. This makes it a suitable choice for systems where high-precision evaluation of user performance is a top priority, particularly in large-scale deployments with sufficient computational resources.
However, the GIA GMDH model demonstrates significant advantages in terms of model simplicity, interpretability, and transparency. Its analytical form allows managers and analysts to explicitly understand how individual behavioral indicators influence performance scores. The reduced number of parameters and faster training time also make it well-suited for systems operating in real-time or resource-constrained environments.
Thus, the selection between these two models should be guided by the specific requirements of the application domain. If maximum accuracy and scalability are essential, GBDT is recommended. In contrast, if model explainability, control over behavior-to-score mapping, or ease of integration is prioritized, GIA GMDH presents a more appropriate solution.
This research explored the integration of intelligent analytics into a web-based project management system with the goal of enhancing individual performance assessment through datadriven methods. Two machine learning algorithms GBDT and the GIA-GMDH were investigated and compared to identify the most appropriate approach for integration into the system.
The comparative analysis revealed that while GBDT offered slightly higher predictive accuracy, GIA GMDH demonstrated greater transparency, interpretability, and analytical compactness. These qualities are essential for real-time feedback and managerial control within the context of project execution. The GIA GMDH model also provided an explicit mathematical formula that enables the direct calculation of performance scores based on a concise set of behavioral indicators, such as task volume, communication, timing, and complexity.
Based on these findings, the GIA-GMDH approach was selected as the preferred model for implementation within the intelligent layer of the project management system. The results validate the potential of combining structured behavioral data with interpretable machine learning models to enhance transparency and decision support in collaborative work environments. However, the study also has limitations. The dataset was based on a controlled pilot environment and may not fully reflect the complexity and variability of real-world project teams.
Future work should explore scaling the approach across diverse organizations, incorporating richer behavioral dynamics, and integrating context-aware or emotion-sensitive inputs. Enhancements such as fuzzy logic or hybrid inference could further improve the system’s robustness under uncertainty and incomplete data.