HURANIA: Leveraging AI to Personalize Workforce Training in Human-Centric Industries

HURANIA: Leveraging AI to Personalize Workforce Training in Human-Centric Industries Chiara Masiero∗, Massimiliano Conte∗, Mariagrazia Colotti†, Massimo Gregori†, Andrea De Clemente† ∗Statwolf Srl [email protected] †LFoundry Srl [email protected] Abstract—We present HURANIA — HUman Resources Artificial iNtellIgence Approaches — an appliedAI system that personalizes corporate training at scale. The solution integrates heterogeneous HR and LMS sources, computes semantic similarity across workers and courses, and deploys a hybrid recommender system (implicit collaborative filtering and a cold-start classifier) within a production-grade MLOps stack. Implemented on the Statwolf analytics platform, HURANIA demonstrates how human-centric AI can automate training assignment, increase catalog coverage, and support continuous upskilling in industrial environments. Index Terms—Personalized learning, recommender systems, embeddings, MLOps, HR analytics, Industry 5.0 I. Introduction As organizations move toward human-centric and resilient Industry 5.0 paradigms, upskilling must adapt dynamically to evolving roles and competencies. HURANIA introduces an end-to-end AI layer for HR operations that: (i) unifies HR and learning data, (ii) represents courses and worker careers with text embeddings, and (iii) recommends relevant training while ensuring production reliability through MLOps. The project is led by LFoundry with Statwolf as the analytics and deployment partner. II. Data Integration and Foundation A. Sources and scope Two key enterprise systems provide the project’s data foundation: • GEPE [ 1 ]: an HR management suite that provides detailed information about employees, including demographics, positions, educational background, organizational units, and employment history. • Totara [ 2 ]: a Learning Management System (LMS) that tracks training programs, certifications, learning events, completion rates, and assigned audiences. A custom integration pipeline automatically extracts reports from Totara in CSV format, stores them in a dedicated S3 blob container, and imports them into the Statwolf platform [ 3 ] through native connectors. The integration process includes cleaning, normalization, and the creation of logical relationships between heterogeneous entities such as employees, job positions, and training programs. This process ensures that all HR and learning data are available for analytics and machine learning in a unified schema. B. Platform capabilities The Hurania solution is built upon the Statwolf data analytics platform, a cloud-enabled framework for end-toend data orchestration, transformation, and model deployment. Its modular architecture includes a web-based client for visualization, a core backend executing business logic, and independent worker units performing computationally expensive tasks asynchronously. Statwolf provides webbased dashboards, developer APIs, and scalable processing (e.g., Hadoop/Spark; ClickHouse/PostgreSQL backends). Data are ingested via native connectors and exposed for downstream ML and reporting. The harmonized dataset enables consistent features for both similarity analytics and recommendation. III. Representation Learning A. Worker embeddings The first analytical step consists of characterizing the workforce to quantify similarity between employees based on their career trajectories. Each employee is represented by a sequence of job titles with temporal information. These textual job titles are encoded into vector embeddings using the text-embedding-ada-002 model, which transforms natural language descriptions into numerical vectors capturing semantic meaning. For each employee u , the system aggregates the embeddings of all positions held, weighting more recent or current roles more heavily: vu=1 W n X i=1 wi·ei, where e i is the embedding of the i -th job title and wi ’s are temporal decay weights summing up to W. Similarity between employees is computed as the cosine similarity of their embedding vectors: sim(u, v) = vu·vv ∥vu∥∥vv∥. This quantitative framework enables the identification of employees with comparable experience profiles, which is fundamental for clustering, benchmarking, and personalized course recommendation. B. Course embeddings Each course in the LMS is similarly represented by a vector embedding derived from its title and description. We encode each course title using general-purpose text embeddings ( text-embedding-ada-002 ). Cosine similarity in the embedding space supports “related courses” exploration and catalog curation. IV. Personalized Recommendation A. Implicit collaborative filtering Given extensive implicit feedback (user–course interactions), the core recommender system is based on an implicit collaborative filtering model using the Alternating Least Squares (ALS) algorithm [ 4 ]. An interaction matrix R∈RU×C encodes whether user u has participated in course c . Entries with Ru,c = 1 represent implicit positive feedback, while missing entries are treated as unobserved interactions. The ALS algorithm factorizes R into latent matrices P (for users) and Q(for courses), minimizing: min P,Q X u,c (Ru,c −P⊤ uQc)2+λ(∥Pu∥2+∥Qc∥2). The scalar product P⊤ uQc estimates the implicit affinity between the user and the course. Recommendations are scored via dot-products in the latent space; already completed courses are excluded at inference time. B. Cold-start classification For new users without an interaction history, the System uses a classification-based approach. Each user–course pair is labeled as positive (course attended) or negative (not attended). A CatBoost model predicts the probability of course engagement using user and course embeddings as features. Given a new user, the system ranks available courses by predicted probability, excluding already completed items. The best performance was achieved by the CatBoost classifier [5], as summarized in Table I. TABLE I CatBoost classification performance Class Precision Recall F1-score Accuracy 0 92% 95% 0.93 – 1 85% 76% 0.80 – Overall – – – 90% C. Serving logic At inference: 1) If collaborative factors exist for a user, serve topk ALS recommendations. 2) For new users (or sparse histories), fall back to the classifier’s propensity ranking. 3) Optionally blend both scores where appropriate; always filter out courses already completed. V. Productive Deployment A. Pipelines and orchestration To ensure scalability and maintainability, Hurania integrates advanced MLOps practices directly within the Statwolf platform. Each training or inference run triggers an automated pipeline that: •logs all training metadata and validation metrics, • stores the model version in a centralized Vault with unique ID and timestamp, • updates monitoring dashboards with real-time performance indicators, • activates alerting mechanisms in case of model drift or anomalies. The recommender system’s pipeline is executed weekly, leveraging automated data integration updates. Each new model is validated, registered, and benchmarked against previous versions before release to production. All predictions are versioned, enabling full traceability between models, datasets, and generated recommendations. This MLOps orchestration ensures reliable, repeatable, and auditable AI operations, aligned with industrial-grade deployment standards. B. Monitoring and auditability Operational dashboards surface model versions, executions, and performance indicators (hit ratio, coverage). Alerts are triggered upon data drift or degradation, facilitating rapid rollback or retraining. The approach ensures traceability from any prediction to the originating model version and data slice. VI. Results The performance of the recommendation system was evaluated through an online assessment involving human experts from the HR and Training teams. The evaluation aimed to verify both the relevance and the novelty of the recommended courses under realistic production-like conditions. A. Evaluation procedure To ensure representativeness, the evaluation covered a heterogeneous set of 70 employees, distributed across 5 corporate departments and 39 organizational units, representing a wide range of roles and skill profiles. For each user, up to five course recommendations were generated, explicitly excluding mandatory certification courses (which are managed separately by the existing training system). The resulting 329 recommendations were independently reviewed by domain experts, who judged whether each suggested course was relevant to the user’s professional development. Experts also identified cases in which a suggestion was not immediately obvious yet still valuable—capturing instances of useful novelty beyond standard training paths. This qualitative feedback served as the ground truth for the computation of the quantitative metrics reported below. TABLE II Quantitative evaluation of recommendation performance Metric Value Hit Ratio@5 61.43% Hit Ratio Novel@5 57.14% B. Quantitative results Two main performance indicators were used: Hit Ratio@5 and Hit Ratio Novel@5. • Hit Ratio@5 quantifies the percentage of users for whom at least one relevant course appeared among the top–5 recommendations, measuring the overall correctness of the system’s output. • Hit Ratio Novel@5 measures the percentage of users for whom at least one relevant recommendation concerned a course offered by a department different from the user’s own, thus evaluating the system’s ability to surface cross-domain and non-trivial learning opportunities. Table II summarizes the quantitative outcomes, showing that the system correctly identified relevant courses for more than 60% of users and was able to propose novel, cross-department suggestions in over half of the evaluated cases. C. Qualitative impact The system demonstrated a strong ability to generate relevant and sometimes unexpected training suggestions, with more than half of users receiving at least one valuable course outside their primary domain. In operational terms, the tool can serve as a decision-support system for supervisors, helping them enrich employee development plans while avoiding training overload. Similarity-based visualizations and embedding-level explanations provide transparency, while version-controlled MLOps pipelines ensure traceability and trust throughout the model lifecycle. VII. Conclusion and Future Work The Hurania system acts as a bridge between human resource management and modern AI infrastructure, enabling a human-centric digital transition consistent with the Industry 5.0 vision. In particular, Hurania demonstrates how applied AI can enhance workforce development by combining: (i) robust data integration, (ii) semantic representations of workers and courses, (iii) hybrid recommendation, and (iv) production-grade MLOps. Future work includes enriching inputs with course descriptions and skill taxonomies, incorporating performance indicators, and extending recommendations toward career-path planning and reskilling at enterprise scale. Acknowledgment This work was supported by the iNEST project under MUR – Missione 4, Componente 2, Investimento 1.5 – Creazione e rafforzamento di “Ecosistemi dell’innovazione”, costruzione di “leader territoriali di R&S” – del PNRR funded by the European Union - NextGenerationEU with grant agreement no. ECS00000043. References [1] Data Management HR Products [2] Totara Learning Management System [3] Statwolf Platform Documentation [4] Hu, Yifan, Yehuda Koren, and Chris Volinsky. "Collaborative filtering for implicit feedback datasets." 2008 Eighth IEEE international conference on data mining. Ieee, 2008. [5] Prokhorenkova, Liudmila, et al. "CatBoost: unbiased boosting with categorical features." Advances in neural information processing systems 31 (2018).