The Science Behind Expertini ATS for hiring teams in Toulouges, OccitanieFrance
A practical guide for recruiters working in Toulouges. View without a localization
The Science Behind Expertini ATS
Where recruitment technology is heading, what Expertini has published about its own methods, and how independent researchers have engaged with that work — stated plainly.
Recruitment software is in the middle of a genuine scientific turn. Research on AI-driven hiring commonly periodises it: traditional machine learning (keyword and statistical matching, roughly 2010–2016), deep learning and word embeddings (2016–2022), large language models (2022–2024), and an agentic era from 2024 onward — arriving together with real regulation, from the EU AI Act to New York City's Local Law 144, which requires bias audits of automated employment decision tools, and a broad shift toward human-in-the-loop oversight as the professional norm.
Expertini's position in that landscape is specific and deliberately conservative: AI is used for what AI is genuinely good at — reading documents and extracting meaning — while the score itself comes from a fixed, published mathematical formula. Same inputs, same score, every time, with dimension-level evidence a human can inspect. This page lays out what we've published about that method, and — separately and without embellishment — where independent researchers have engaged with it.
On this page
01What Expertini has published
Three papers document the method's evolution. A 2024 SSRN paper described the original Resume Score™ approach — cosine-similarity scoring of CVs against job descriptions with NLP-based parsing. A November 2024 TechRxiv preprint, "Leveraging Mathematical and Artificial Intelligence for Automated Resume Screening," formalised the Candidate Match Score: a weighted multi-criteria formula (CMS = Σ(skill score × requirement importance) / Σ importance) and the experimental case for importance-weighted over naive averaging. The 2026 study, "From Stochastic to Deterministic," describes the current hybrid architecture — an AI layer for semantic extraction bounded by the deterministic CMS formula — including its bias-mitigation pipeline (PII stripping, bounded attributes, description-derived weights) and, importantly, a section of acknowledged limitations: probabilistic weight inference, residual training-data bias, and the explicit statement that a CMS score is a screening aid, never a hiring decision.
02Independent engagement with the work
Two independent publications have engaged with Expertini's platform or research — we state them precisely rather than inflate them. A 2026 master's thesis at Luleå University of Technology (Sweden), comparing agent-based, semantic, and traditional models for CV–job matching, cites Expertini's 2026 semantic-recruitment study in its related work — noting the study's finding that semantic models outperform keyword-based ATS matching and can make screening more inclusive for candidates who describe the same skill in different words. Separately, a 2025 peer-reviewed survey in the NIU Journal of Social Sciences (Adeyosola, "Artificial Intelligence and Recruitment Practices in Nigeria," 11(4), 295–300), studying 540 respondents across five Nigerian states, identifies the Expertini platform among the AI applications recruiters actively use for CV scrutiny, resume parsing, candidate shortlisting, matching, and skills validation. A thesis citation and a survey mention are exactly that — early, independent engagement, not endorsements — and we'd rather report them accurately than dress them up.
03Where the wider software world is going — and where we sit
Modern business software has converged on a few norms: systems talk to each other (CRM, ERP, HRIS, and accounting platforms exchange records rather than trapping them), decisions leave an audit trail, and AI is governed rather than merely deployed. Expertini has adopted all three: the connector layer syncs clients, contacts, and hires two-way with Salesforce, Dynamics 365, HubSpot, Odoo, ERPNext, BambooHR, Deel, SAP SuccessFactors, Xero, QuickBooks, and Zoho; every consequential action lands in an attributed audit trail; and every AI-touched decision is reconstructable because the scoring layer is deterministic. The agentic era's open question — how much autonomy hiring systems should have — we answer conservatively: tools inform, humans decide.
04Why determinism matters more as regulation arrives
Bias-audit laws like NYC Local Law 144 and the transparency obligations taking shape under the EU AI Act share one practical demand: you must be able to explain and reproduce what your system did. A stochastic scorer — one that can give the same CV a different score on a different day — cannot honestly meet that demand. A published formula can: every CMS score decomposes into its dimension scores and weights, and re-running the same inputs yields the same result. That's not a marketing preference; it's the property that makes an automated screen defensible when a rejected candidate, a regulator, or your own counsel asks why.
05The honest limits
Published methodology doesn't mean solved problem. Semantic extraction still runs on large language models trained on human text, which encodes human bias; our anonymisation and bounding reduce the pathways, they don't eliminate them. Candidates from professional cultures that quantify achievements still hold a residual advantage. And fair scoring doesn't guarantee fair hiring — shortlisting decisions remain human, which is by design, but it means outcomes depend on the people using the tool. These limits are stated in our own 2026 paper, and repeating them here is part of the point: an ethical screening product should be easiest to criticise from its own documentation.
Frequently asked questions
Is the research peer-reviewed?⌄
Does 'deterministic' mean no AI is involved?⌄
Can I check the methodology myself?⌄
Does any of this replace human judgement?⌄
At a glance
- AI reads; a published formula scores — reproducible by design
- 3 methodology papers + 2 independent academic engagements, reported precisely
- Bias mitigation: PII stripping, bounded attributes, description-derived weights
- Built for the bias-audit era (EU AI Act, NYC Local Law 144)
- CRM/ERP/HRIS interoperability adopted as a design norm
- Limitations documented in our own publications
See the science behind expertini ats on your own hiring.
Bring a real job description to a 30-minute demo — free trial included.
Book a demo