AI Framework Reads Job Postings to Expose Gaps in University Curricula

The disconnect between higher education curricula and the evolving requirements of the global labor market remains one of the most discussed challenges in the twenty-first-century knowledge economy. In the United States, projections by the Bureau of Labor Statistics suggest that significant numbers of positions in data-driven and technology-oriented sectors may remain unfilled in part due to shortages of qualified talent. Such shortages are commonly attributed not only to insufficient graduate output but also to a structural lag between curriculum design and the pace of industrial change.

As artificial intelligence, automation, and data science continue to reshape occupational profiles, higher education institutions face increasing pressure to update their academic programmes. Curriculum revision in many institutional settings depends on multi-year administrative cycles supported by manual surveys, expert committees, and accreditation reviews. By the time new content reaches course syllabi, employer demand may have shifted further.

Universities also tend to rely on relatively narrow data sources, such as alumni feedback or employer questionnaires, rather than continuous, data-driven analyses of labor-market trends. These constraints limit institutional responsiveness to emerging competency areas in AI, cybersecurity, sustainability, and data analytics. Recent advances in natural language processing, sentence embeddings, and Retrieval-Augmented Generation make it more feasible to analyse labor-market text at scale, extract recurring skill signals, and compare them to academic syllabi.

When applied responsibly, these techniques can complement, rather than replace, expert curriculum review by structuring market evidence in a form that supports targeted faculty decision-making. Within this context, a new study introduces CurriculumCatalyst, an open-source prototype framework intended to support periodic, evidence-grounded alignment between higher education programmes and labor-market signals. The framework retrieves textual evidence from public job postings and employment databases, applies an NLP pipeline to extract competency expressions, and compares those expressions to skill signals extracted from academic syllabi.

Where coverage gaps are detected, a multi-level RAG component drafts curriculum artifacts, including module outlines, learning objectives, reading lists, and assessment prompts, that remain consistent with common institutional constraints such as prerequisite sequencing and credit-hour structure. Outputs are surfaced through a Streamlit-based dashboard intended to support, not replace, faculty and curriculum-committee review. The present work contributes a framework and a prototype implementation, accompanied by a preliminary, exploratory evaluation on a pilot corpus.

It does not establish pedagogical validity of the generated artifacts, expert acceptability, downstream educational impact, or comparative superiority over simpler approaches such as keyword matching or TF-IDF. These are identified as future work and are addressed throughout the paper as anticipated capabilities rather than demonstrated outcomes. The early prototype run surfaced coverage gaps above 20 percent in selected technology-oriented topics, including AI programming, cloud computing, and data analytics; the authors treat that figure as a descriptive, prototype-level indicator and not as a population-level estimate.

CurriculumCatalyst operates on two textual corpora that are constructed and refreshed independently. The labor-market corpus is built from publicly accessible job postings and public employment databases, treated as text sources reflecting current employer-stated competency requirements. The academic corpus is built from institutional programme documents and course syllabi, which together describe both explicit learning outcomes and implicit skill signals embedded in topic coverage, project descriptions, assessment language, and reading lists.

For the present pilot, the labor-market corpus consisted of job postings collected within a contemporaneous time window covering technology-oriented occupations such as software development, data analytics, and cloud and AI engineering; the academic corpus consisted of syllabi and programme documents drawn from technology-oriented undergraduate programmes. Documents were retained when they contained either an explicit job-title and required-skills section on the labor-market side or recognisable course-level descriptors such as a title, learning outcomes, or topic list on the academic side. Documents that were too short to support reliable extraction, duplicates identified by URL or near-duplicate hashing, and pages that consisted primarily of navigation or boilerplate were excluded.

The authors do not present the pilot corpus as a probability sample of either employer demand or academic offering; it was assembled to exercise the end-to-end pipeline and to surface concrete examples of detected gaps. After preprocessing, which comprised HTML stripping, language filtering, Unicode normalisation, sentence segmentation, lemmatisation via spaCy, and removal of standard English stop words augmented with a domain-specific stop list, the framework extracts structured skill signals from both corpora using a combination of rule-based and embedding-based methods. Candidate competency phrases are first identified through a lexicon of skill terms compiled from publicly available labor-market taxonomies such as ESCO and O*NET, and through dependency-pattern matching in spaCy that targets noun-phrase constructions associated with task or tool descriptions, for example phrases like experience with X, proficiency in Y, or ability to perform Z.

Each candidate phrase is then encoded into a dense vector using a Sentence-BERT model, a general-purpose sentence encoder in the all-MiniLM family in the present implementation. On the curriculum side, the same encoder is applied to learning outcomes, course descriptions, and topic lists, so that academic and market phrases occupy a shared embedding space. Curriculum phrases that are not articulated as outcomes are still captured as implicit skill signals through this embedding-based extraction, which is one reason for using embeddings rather than relying on string matching alone.

Alignment analytics compares the two skill-signal sets in the shared embedding space. For each labor-market skill phrase, the framework computes its cosine similarity to every curriculum skill phrase, and defines a market phrase as covered when at least one curriculum phrase exceeds a similarity threshold. In the present pilot the threshold was set at 0.65 as a default, selected by inspection of paired examples in a held-aside calibration set drawn from the same corpora; the authors additionally report sensitivity by re-running the analysis at thresholds of 0.55, 0.60, and 0.70 so that gap estimates are not interpreted as point-true values.

The framework also de-duplicates near-identical market phrases prior to scoring, using a cosine similarity cutoff of 0.90, so that frequently repeated phrases do not inflate the apparent demand for a single skill. The skill-coverage gap for a given programme or course set is then operationalised as the proportion of distinct, weighted in-demand market skills that lack a covered counterpart in the curriculum representation. Weighting is by document frequency in the labor-market corpus, capped to limit the influence of templated repostings.

The resulting metric is descriptive: it summarises an embedding-based alignment over the specific corpora and threshold used, rather than a population-level estimate of curriculum quality. A central methodological component is a multi-level Retrieval-Augmented Generation mechanism that separates evidence retrieval from curriculum content generation. In the retrieval stage, the system indexes labor-market documents at the chunk level in a vector store and, for each gap identified, retrieves the top five chunks most semantically related to the missing or underrepresented skill.

These retrieved chunks anchor recommendations in observable market evidence rather than in generic content templates. In the generation stage, a large language model is prompted with the gap description, the retrieved evidence chunks, a structured set of programme constraints including existing prerequisites, credit-hour budget, target programme level, and accreditation context where supplied, and a fixed output schema asking for a module title, learning objectives expressed in measurable verbs, suggested topics, recommended readings, indicative assessment items, and recommended placement within the existing programme sequence. Prompt templates are kept in the open-source repository so that they can be inspected, replaced, or domain-adapted.

Generated outputs are written to the dashboard for review rather than committed automatically to any institutional record. To make the pipeline interpretable, the authors present one illustrative trace from the pilot run. In the labor-market corpus, the competency of containerisation and orchestration with Docker and Kubernetes appeared in multiple postings for cloud and DevOps engineering roles, supported by phrases such as build and deploy containerised services, manage Kubernetes clusters in production, and configure CI/CD pipelines for container workloads.

In the academic corpus for the same pilot, the closest curriculum phrases were virtualisation and operating-system isolation from a systems-programming course, introduction to cloud platforms from an electives module, and software deployment fundamentals from a software-engineering course. At the default threshold of 0.65, none of the curriculum phrases met the coverage threshold relative to the market expression of the competency, and the framework therefore recorded the competency as not covered. At a threshold of 0.55 the introduction to cloud platforms phrase met the threshold, illustrating the threshold-sensitivity property reported in the study.

The retrieval stage subsequently surfaced five labor-market chunks describing the competency in employer language. The generation stage produced a draft module titled Cloud-Native Containerisation and Orchestration with five learning objectives expressed in measurable terms, a topic list, indicative readings drawn from open educational resources, two assessment items including a hands-on lab and a design exercise, and a recommended placement after the existing software-engineering course as a 3-credit elective. Provenance metadata in the dashboard links each retrieved chunk to its source posting, and the curriculum-side comparison links to the specific syllabus spans evaluated.

The authors present this example as an illustration of the pipeline's outputs and traceability, not as a validated curriculum recommendation. Several risks are inherent to the present design and require explicit acknowledgement. First, job postings reflect employer-stated demand on the platforms where they are observed, not the full set of professional competencies needed for sustained practice; they may also encode geographic and platform skews and may inflate certain keywords for search purposes.

Second, the embedding-based matching can produce false positives, semantically near phrases that are pedagogically distinct, for example machine learning theory versus applied machine learning, and false negatives, genuinely covered competencies missed because the syllabus phrasing is unusual. Third, curriculum coverage is sensitive to syllabus completeness; programmes whose syllabi under-document topics will appear to have larger gaps than programmes whose syllabi are richly written, even when the underlying instruction is similar. The authors mitigate these risks where feasible by reporting sensitivity to the similarity threshold, by carrying provenance metadata so that any flagged gap can be traced back to a specific market document and a specific curriculum span, and by presenting outputs as drafts subject to human review rather than as decisions.

They do not claim that these mitigations are sufficient. In particular, fairness-aware analysis across regions, languages, and demographic axes, and systematic calibration of false-positive and false-negative rates against expert-labeled data, remain open work. The prototype results indicate that an end-to-end pipeline linking labor-market retrieval, semantic alignment, and constraint-aware generation is technically feasible and produces interpretable outputs with traceable provenance.

They do not indicate that such a pipeline produces pedagogically valid curriculum modules, that its detected gaps would survive expert scrutiny, or that it outperforms simpler analytic methods. Readers should treat the framework as a structured way of organising labor-market evidence in a form that is reviewable by faculty, not as a substitute for academic judgment. Priorities for future research follow directly from the limits of the present evaluation: a head-to-head benchmark against keyword, TF-IDF, and non-RAG NLP pipelines using expert-labeled gap data with precision, recall, and F1 reporting; a multi-institution, multi-discipline evaluation including humanities and professional programmes to test cross-domain transfer; faculty and curriculum-committee review studies with inter-rater reliability and usability instruments; longitudinal evaluation of curriculum changes adopted with framework support; explicit fairness analysis of labor-market sources including geographic, sectoral, and language coverage; and implementation of streaming ingestion and drift-aware re-retrieval to make the system genuinely continuous in operation rather than only in design intent.

Within these scoped claims, CurriculumCatalyst contributes a structured, extensible pathway for translating labor-market evidence into reviewable curriculum diagnostics and drafts, and a transparent foundation on which such broader empirical studies can be built.

Subject of Research: An AI-based framework for continuous alignment between higher education and industry skill needs

Article Title: An AI-based framework for continuous alignment between higher education and industry skill needs

Article References: Sunday, D., Jingo, F., & Kimbugwe, J. (2026). An AI-based framework for continuous alignment between higher education and industry skill needs. Discover Education, 5(1), Article 893. https://doi.org/10.1007/s44217-026-02083-8

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02083-8

Keywords: AI-based, framework, continuous, alignment, between, higher, education, industry, skill, needs, scientific research