Taxonomy Unification AI for Skills Normalization

The client is a NASDAQ-listed professional-services firm and one of the longest-running partnerships in the portfolio — a multi-year relationship that started with internal engineering services and expanded into a delivery channel serving the partner's external clients across consumer goods, industrial manufacturing, and professional-services markets. The partnership has placed thirty-plus engineers inside the staffing firm's own teams and more than 180 across the partner's external client work, with thirteen distinct projects.

The staffing firm's recruiting data was spread across systems, job descriptions, candidate profiles, resumes, and internal recruiting workflows, but the language was inconsistent. The same skill could be described in multiple ways. The same role title could mean different things depending on the employer, industry, seniority level, or business context. A "developer," "software engineer," "full-stack engineer," and "application engineer" could overlap heavily in one context and mean very different profiles in another.

This created a structural problem for AI matching. Even strong matching models produce weak results when the underlying language is fragmented. If skills are not normalized, role families are not consistently represented, and job descriptions are not converted into a common taxonomy, the system cannot reliably compare one job to another or one candidate to one job. The staffing firm needed more than a data-cleaning exercise. It needed an AI-driven taxonomy layer that could understand recruiting language, group related skills and jobs, normalize messy descriptions, and create a consistent representation of demand across the business.

Taller designed and built an AI taxonomy engine for recruiting data, combining NLP, unsupervised learning, supervised classification, deep learning, and custom data-engineering pipelines. The goal was to create a proprietary recruiting taxonomy that could represent jobs, skills, and role families in a consistent way across the staffing firm's data.

The first layer focused on skill detection. Taller built NLP pipelines to extract skills from job descriptions and candidate profiles, identifying both explicit skills and related terms that appeared in different wording. This allowed the system to detect that two descriptions could be asking for the same underlying capability even if the language was different.

The second layer focused on grouping and normalization. Taller used unsupervised learning techniques, including clustering and K-nearest-neighbor-style similarity analysis, to group related jobs and related skills. These models helped identify natural clusters in the data: groups of skills that commonly appeared together, job descriptions that belonged to the same functional area, and role families that were semantically similar even when their titles were different.

The third layer was taxonomy creation. Taller used the output of the AI models to build a normalized skills and jobs taxonomy tailored to the staffing firm's recruiting business. Instead of relying only on generic public taxonomies, the system learned from the staffing firm's own market data and recruiting patterns. Job descriptions were transformed into structured representations using this taxonomy, allowing them to be compared, searched, classified, and matched more accurately.

The fourth layer connected the taxonomy to recruiting intelligence. Once jobs and skills were normalized, the system could classify roles by business area, such as technology or finance, detect demand patterns, improve job-order matching, and support candidate probability scoring. The taxonomy became a service layer that other recruiting workflows could consume, turning unstructured job language into machine-readable recruiting intelligence.

This work was especially meaningful because it was built before the current LLM wave. Taller did not rely on generic language models to infer everything automatically. The team built the intelligence through custom NLP, machine-learning research, clustering, neural-network modeling, labeled and unlabeled datasets, domain-specific feature engineering, and a taxonomy designed around the way staffing and recruiting actually work.

The taxonomy engine reduced data discrepancies by fifty percent, improved candidate-matching accuracy by thirty percent, and increased recruiting efficiency by forty percent. More importantly, it gave the staffing firm a reusable AI foundation for matching, classification, job analysis, and recruiting intelligence. Jobs and skills that had previously been fragmented across inconsistent language could now be represented in a normalized structure that downstream systems could use.

Taxonomy quality is the foundation for AI in recruiting. Without it, matching models operate on noisy language and produce low-confidence results. Taller's work gave the staffing firm a normalized recruiting intelligence layer: a system that could understand skills, group similar jobs, classify demand, and represent recruiting data in a way that AI models could actually use.

Strategically, this was one of the precursors to Echo. It proved that the core problem in staffing AI is not only automation, but language normalization: turning messy job descriptions, inconsistent skills, and fragmented recruiting data into a structured intelligence layer. That same thesis later became central to Echo's matching and sourcing capabilities — the ability to understand jobs, candidates, skills, and market demand through a taxonomy that reflects how recruiting actually works.