Abstract:
Modern recruitment faces a critical challenge: state-of-the-art semantic search engines, while powerful, operate as opaque ”black boxes” that lack the explainability and specificity required by professional recruiters. This project addresses this challenge through a dual-track approach. First, it details the development of a high-performance Industrial Baseline Engine, a system delivered to the host company that uses Google Gemini and the FAISS library to fulfill all requirements of a modern semantic matching solution.
Second, using the insights gained from the baseline’s inherent limitations, this thesis introduces an advanced research framework. This framework is named MAALEM, an acronym for Multi-Axis Agile Learning Engine for Matching. The name, which translates to ”master” or ”expert” in Arabic and is also the surname of the author, reflects the project’s core ambition: to achieve a masterful and nuanced understanding of person-job fit. The MAALEM framework’s core innovation is its deconstructed architecture, which moves beyond a single similarity score to evaluate candidates along multiple, transparent, and interpretable dimensions. This design prioritizes justified reasoning and user control, transforming the AI from an opaque filter into a genuine decision-support partner. To validate this new paradigm, a comprehensive suite of three novel benchmarks was created, including an adversarial test designed to probe model intelligence and resilience. Empirical results demonstrate that while the Industrial Baseline is a powerful tool, the MAALEM framework significantly outperforms it and all other standard models, especially under adversarial conditions. It proved uniquely effective at rejecting deceptive candidates while simultaneously identifying non-obvious ”hidden gem” profiles. MAALEM therefore represents a new vision for recruitment AI one that is not only more accurate and robust but is also fundamentally more trustworthy and strategically aligned with the needs of its expert users.
