• Gaspar Jobs Expertini profile avatar
    • Expertini CMS Engine  ·  MCDA + Gemini AI

    Introducing the Expertini Candidate
    Match Score     (CMS)

    Deterministic · Bias-Free · Auditable

    The Expertini Candidate Match Score brings the same scientific discipline that powers ERIS — the researcher impact framework — to the world of talent acquisition. Upload any CV, paste a Job Description, and our pipeline strips all personal identifiers before Gemini AI extracts weighted competency dimensions. The Expertini CMS formula then scores the match deterministically — producing a reproducible, explainable result free from AI hallucination and demographic bias.

    ⚡ Score a Candidate Now → Read the Formula ↓ FAQ ↓
    CMS = Σ (CSSi × JRISi) / Σ JRISi

    What is the Expertini Candidate Match Score?

    The Expertini Candidate Match Score (CMS) is a deterministic, multi-criteria weighted scoring framework that measures how well a candidate's documented professional experience aligns with a specific job description. It is calculated by combining two inputs per competency dimension: a Candidate Skill Score (CSSi) — how well the candidate evidences the competency — and a Job Requirement Importance Score (JRISi) — how critical that competency is to the role. These are combined using a weighted average bounded to produce a final score between 0 and 100.

    CMS is not intended to replace the judgement of a skilled recruiter or hiring manager. It is a precision screening layer designed to surface well-matched candidates who might be discarded by keyword filters, and to mathematically surface critical gaps that unconstrained AI would overlook. The score is fully auditable — every dimension, every weight, every calculation is visible. Use CMS as your first filter, not your final verdict.

    "No single number captures the full complexity of a candidate's potential. A multi-criteria, weighted score offers a more balanced and defensible measure than any keyword count or unconstrained AI opinion taken in isolation — but the score is still a starting point, not a hiring decision."
    0–100
    CMS Scale

    A single number. A bounded story.

    CMS combines AI-extracted competency evidence with mathematically inferred job importance weights into one reproducible score — capped at 100 to prevent any single competency from dominating and to ensure fair comparison across roles, industries, and candidate backgrounds.

    🪐 Inspired by ERIS (Planet) The Expertini Researcher Impact Score (ERIS) was built on the same principle: strip the noise, bound the mathematics, and let the data speak without institutional bias. CMS applies that philosophy to hiring — reaching beyond surface-level keyword matching to evaluate genuine competency alignment, wherever the candidate comes from and however they write.

    The Two Core Components of CMS

    Each CMS score is built from two values extracted for every competency dimension. The design separates what the candidate brings from what the role demands — a distinction that flat keyword scoring has never made.

    Candidate Input · 0–100 per dimension
    🎯 CSS — Candidate Skill Score
    The CSS measures how well the candidate's anonymised CV evidences a specific competency dimension. Gemini AI reads the professional content — achievements, timelines, metrics, leadership signals — and assigns a score between 0 and 100. A missing mandatory qualification scores CSS = 0, creating the mathematical hard-stop the role demands. Elegant writing cannot inflate CSS; only documented evidence does.
    Role Input · 0–100 per dimension
    ⚖️ JRIS — Job Requirement Importance Score
    The JRIS captures how critical each competency is to the employing organisation — inferred automatically by Gemini AI from the language of the job description. "Must possess" and "non-negotiable" map to JRIS = 100. "Preferred experience" maps to 60–75. "Familiarity is a plus" maps to 30–50. No manual slider configuration is required from the recruiter — the JD text drives the weights automatically.

    The mathematical consequence of this two-component structure is significant. When a candidate is missing a JRIS = 100 dimension — a mandatory requirement — that zero-multiplication effect adds 100 to the denominator while contributing nothing to the numerator. The final score drops materially regardless of how strong the candidate is in every other dimension. This is not a penalty imposed by the system; it is the mathematical truth the recruiter already knew made explicit and reproducible.

    The CMS Formula — Defined

    The CMS formula is a weighted mean, fully reproducible from Gemini's structured output. Each competency dimension is scored independently and weighted by its importance to the role. The weighted scores are summed and normalised by the total importance weight. Formally:

    // Expertini Candidate Match Score (CMS)
    // Reference: Syed et al. (2026) — From Stochastic to Deterministic

    CMS = [ Σ (CSSi × JRISi) ] / [ Σ JRISi ]

    // Where:
    i = a competency dimension extracted from the job description
    CSSi = Candidate Skill Score for dimension i [0, 100]
    JRISi = Job Requirement Importance Score for i [0, 100]

    // Note: Because CSS and JRIS are both 0–100, the ratio naturally
    // lands on a 0–100 scale. No × 100 multiplier is needed or correct.

    // Worked example (Owen Wright / Classified Python Engineer role):
    Numerator = (100×100) + (90×100) + (85×95) + (80×90) + (50×90) + (60×60) + (100×0)
    = 42,375
    Denominator = 100 + 90 + 85 + 80 + 50 + 60 + 100 = 565
    CMS = 42,375 / 565 = 74.96

    The 74.96% result tells a precise story. Owen Wright is a world-class Python engineer — his technical scores are near-perfect across six of seven dimensions. But he has no active security clearance, and that dimension carries JRIS = 100 because the job description states it is non-negotiable. The zero-multiplication effect suppresses the final score to 74.96%, correctly positioning Owen for talent pipeline consideration while making clear he cannot be immediately placed in a classified environment. A pure LLM might have scored him at 92% and missed the compliance flag entirely. A keyword filter would have scored him at 36% and discarded him without a human ever reading his profile.

    "The formula permits Gemini to interpret meaning while forcing the final hiring signal to remain grounded in real parameters and reproducible mathematics. AI reads the text. The formula governs the outcome."
    — Syed, Habeebi & Habibi, 2026

    Calculate a Candidate Match Score

    Upload the candidate CV and paste the job description. Personal identifiers are stripped automatically before Gemini AI sees any content. Results are calculated using the CMS formula and can be exported as a full PDF report.

    📄 Upload Candidate CV

    PDF or DOCX. Names, emails, phone numbers, addresses, age proxies, nationality and gender markers are removed before AI processing. Only anonymised professional competency content is evaluated.

    🛡️

    Privacy-First Pipeline All personal identifiers are stripped server-side via a regex anonymisation layer before the CV text reaches Gemini AI. The model only evaluates professional skills, metrics, and achievements.

    📂
    Drop CV here or click to browse
    PDF, DOC, DOCX · max 10 MB

    📋 Job Description

    Paste the full job description. Gemini AI infers competency dimensions and importance weights (JRIS) automatically from your language — no manual configuration required.

    An error occurred.

    — CMS
    —
    —
    CMS = Σ(CSS × JRIS) / Σ JRIS × 100
    Dimensions
    —
    extracted
    Numerator
    —
    Σ CSS×JRIS
    Denominator
    —
    Σ JRIS
    CMS Score
    —
    of 100
    🛡️ Privacy Report

    Personal identifiers stripped before AI processing.

    —
    PII Reduced
    —
    Chars Removed
    ✅
    Bias-Free
    Competency Dimension JRIS CSS Weighted Rationale & Evidence
    Methodology Reference

    Syed, A. H., Habeebi, S. A., & Habibi, S. M. M. (2026). From Stochastic to Deterministic: A Multi-Criteria Decision Analysis Framework for Bounded Semantic Parsing in AI-Driven Recruitment Screening. Expertini Research Department, Washington D.C. / London / Hyderabad. View Paper →

    How CMS Works — Four Stages

    Each stage has a distinct, separated responsibility. AI interprets meaning. Mathematics governs the outcome. No single component is asked to do the other's job.

    📂
    1
    CV Extraction
    PDF or DOCX uploaded. Text extracted server-side using PyPDF2 and docx2txt. No file is stored permanently. Extraction handles both selectable-text PDFs and structured DOCX formats.
    🛡️
    2
    PII Anonymisation
    Names, emails, phone numbers, addresses, age proxies (graduation years), nationality, gender markers, social profile URLs, and postcode patterns are removed via a 10-pattern regex pipeline before any AI processing.
    🤖
    3
    Gemini Semantic Extraction
    Anonymised CV and JD sent to Gemini AI. The model extracts 5–9 competency dimensions from the JD and assigns a CSS (0–100) and JRIS (0–100) to each, based on evidence in the CV and linguistic importance signals in the JD.
    📐
    4
    CMS Formula
    CMS = Σ(CSSi × JRISi) / Σ JRISi × 100. Fully deterministic and reproducible. Missing mandatory requirements score CSS = 0 and mathematically suppress the final score regardless of strength in other areas.

    Why Not Just Ask an AI to Score?

    If you ask an unconstrained large language model — Gemini, GPT-4, Claude, or any other — to "read this CV and rate the candidate out of 100," you will receive a number. That number will be generated confidently. It may be completely wrong in ways that are invisible without an audit trail.

    Large language models suffer from what the CMS paper identifies as "over-excitement" — a tendency to be captured by elegant, metric-dense, confidently-written prose. A CV that reads with corporate polish triggers high linguistic activation regardless of whether the underlying experience matches the role requirements. A required security clearance that is never mentioned in the CV may be overlooked entirely if the rest of the CV is impressive enough. The model's output is stochastic — run the same query twice at different temperatures and you may get different results.

    Evaluation Method Deterministic? Context-Aware? Bias-Resistant? Auditable? Hard Stop on Missing Requirements?
    Keyword / BM25 Filter ✓ Yes ✗ No ✗ No ~ Partial ✗ No
    TF-IDF Cosine Similarity ✓ Yes ~ Partial ✗ No ~ Partial ✗ No
    Pure LLM (unconstrained) ✗ No ✓ Yes ✗ No ✗ No ✗ No
    Expertini CMS (Gemini + MCDA) ✓ Yes ✓ Yes ✓ Yes ✓ Absolute ✓ CSS = 0

    The CMS framework solves this by using Gemini only where it genuinely excels — semantic understanding and contextual extraction — while the deterministic CMS formula handles everything that requires reproducibility, auditability, and hard constraint enforcement. Gemini reads the text. The formula governs the score.

    How CMS is Designed to Be Bias-Free

    The history of automated hiring is also a history of automating discrimination. Amazon's internal ML recruiting tool, abandoned in 2018, taught itself to downgrade CVs containing the word "women's" because it was trained on a decade of predominantly male historical hires. CMS is explicitly designed to resist this through a combination of structural anonymisation and mathematical bounding.

    🚫
    Name Removal
    Candidate names are stripped before AI sees any CV content. Research (Bertrand & Mullainathan, 2004) documents endemic name-based discrimination in hiring — this pipeline eliminates that vector entirely.
    📅
    Age Proxy Removal
    Graduation years and age signals are normalised away. The system calculates competency evidence without exposing when the candidate's career began.
    🌍
    Demographic Stripping
    Nationality, citizenship, marital status, gender markers, religion, and geographic sub-locality are removed. The model evaluates only professional competency signals.
    ✍️
    Style Cannot Inflate Score
    The CMS formula caps every dimension at 100. Confident, polished writing yields zero additional mathematical weight once competency is validated. The algorithm evaluates what the candidate has built, not how they describe it.
    ⚖️
    JD-Derived Weights
    JRIS weights come from the job description — not from profiles of historical hires. This decouples scoring from any bias present in an organisation's past hiring decisions.
    🔍
    Full Auditability
    Every dimension, weight, and score is exposed in the breakdown table. Any candidate or recruiter can see exactly which competency gap drove the result — no black box, no unexplainable output.
    "The problem with most automated screening tools is not that they measure the wrong things — it is that they embed assumptions about which writing styles, institutions, and backgrounds are legitimate. CMS attempts to strip those assumptions out and let the professional evidence speak without demographic ventriloquism."

    CMS is One Measure Among Many

    CMS does not claim to capture everything that matters in a hiring decision. It is a precision screening layer, not a complete assessment of a candidate's value or potential. The following are genuine indicators of fit that CMS cannot measure:

    • 🧠
      Cultural fit and interpersonal dynamics. No scoring system can evaluate whether a candidate will collaborate effectively with a specific team, adapt to an organisation's working style, or communicate well under pressure. Structured interviews and reference checks remain essential.
    • 📈
      Learning trajectory and potential. A candidate at 70% CMS who has been growing rapidly in the right direction may be a better hire than one at 85% CMS who has plateaued. CMS measures documented evidence at a single point in time, not the slope of a career.
    • 🌱
      Unconventional backgrounds. Career changers, self-taught practitioners, and candidates from non-traditional paths may have genuine capability that their CV vocabulary does not yet fully capture. CMS partially addresses this through semantic matching — but professional judgement remains essential for non-linear profiles.
    • ✍️
      CV quality variance by cultural background. Candidates from cultures where metric-dense self-promotion is less normative may produce CVs that contain excellent experience described without the quantified achievement language CMS rewards. This is a known limitation that future calibration work will address.
    • 🤝
      Motivation and alignment. Why a candidate wants this role, how they think about the organisation's mission, and whether they are genuinely committed to the work cannot be read from a CV. That is what interviews are for.
    "Use CMS as your first filter — a fast, transparent read of a candidate's documented competency alignment. Do not use it as your final answer. Hiring decisions are too important to be reduced to a single number, however carefully constructed."
    — Expertini CMS Methodology Notes

    What CMS Relies On — and Its Boundaries

    The CMS framework is transparent about the distinction between what the system evaluates directly and what it accepts from the CV as presented. A score is only as reliable as the evidence it is built on.

    ✓ What CMS Evaluates Directly
    • Semantic proximity between CV content and JD requirements
    • Presence or absence of mandatory qualifications (hard blockers)
    • Documented metrics and quantified achievements in the CV
    • Career timeline signals and seniority indicators
    • Linguistic importance signals in the job description (JRIS inference)
    · What CMS Cannot Independently Verify
    • Whether claimed qualifications, certifications or clearances are genuine
    • Accuracy of stated metrics (e.g. "improved performance by 50%")
    • Employment dates or institutional affiliations listed in the CV
    • Whether experience claimed was individual or collaborative
    • The candidate's current availability or compensation expectations

    Frequently Asked Questions

    Questions we actually get asked — answered without jargon.

    What is a good CMS score? ▾
    There is no universal threshold — a "good" score depends on the role and the candidate pool. As a practical guide: 85+ indicates an exceptional match with strong evidence across all key dimensions. 70–84 is a strong match, likely worth interviewing even if some secondary requirements are unmet. 55–69 is a moderate match — one or two meaningful gaps but a capable overall profile. Below 55 typically indicates significant misalignment or missing core requirements. CMS scores should always be read alongside the dimension breakdown, not in isolation.
    Why does a missing requirement drop the score so dramatically? ▾
    Because that is mathematically what it means for a requirement to be non-negotiable. When a dimension carries JRIS = 100 (a hard requirement) and the candidate scores CSS = 0 (no evidence), the zero-multiplication adds 100 to the denominator while contributing nothing to the numerator. The formula accurately reflects the hiring reality: no matter how qualified the candidate is in every other respect, they cannot fulfil the role as currently defined. This is not a penalty — it is the mathematical truth made visible and reproducible.
    Can I use any LLM instead of Gemini? ▾
    Yes. The CMS framework is model-agnostic by design. The CMS paper explicitly notes that Gemini was selected for empirical validation but that GPT-4o, Claude, Mistral, LLaMA, and other sufficiently capable LLMs can serve as the semantic extraction layer. The scientific contribution of CMS is the deterministic bounding formula, not the upstream model choice. Organisations with strict data residency requirements may prefer on-premise models like Mistral or LLaMA.
    Is the CV stored or shared after scoring? ▾
    No CV content is stored permanently by the Expertini CMS tool. The file is processed server-side for text extraction and PII anonymisation, then the anonymised text is sent to Gemini for analysis. Neither the original file nor the extracted text is persisted to a database after the API response is returned. The only data retained is the structured scoring output that you see on screen and can export as a PDF.
    How many competency dimensions does CMS extract? ▾
    Gemini extracts between 5 and 9 competency dimensions per evaluation, derived from the specific job description provided. The system does not use a fixed template of dimensions — every evaluation is driven by the actual requirements of the role. A technical role will produce dimensions around programming languages, system architecture, and compliance; a commercial role will produce dimensions around sales methodology, CRM proficiency, and communication skills.
    Can CMS be gamed by adding keywords to a CV? ▾
    It is significantly harder to game than keyword-based systems. Gemini evaluates conceptual proximity and looks for evidence signals — documented metrics, career timelines, leadership indicators — not surface keyword density. A CV that simply copies keywords from the job description without backing them with documented professional evidence will not produce a high CSS, because the AI is evaluating the claim against the evidence, not just confirming the term is present. That said, no system is completely immune to manipulation by a sufficiently determined candidate. CMS reduces this risk substantially compared to ATS keyword filters.
    Is CMS the same as the ERIS score? ▾
    No — they are sibling frameworks built on the same design philosophy but for different domains. ERIS measures a researcher's scholarly impact using publications, citations, and h-index. CMS measures a job candidate's professional alignment with a specific role using AI-extracted competency scores and job-requirement importance weights. Both are deterministic, bounded to 0–100, designed to be bias-resistant, and fully auditable. CMS was explicitly inspired by ERIS — the same principle of bounding AI outputs with mathematical governance, applied to hiring.
    Where can I read the full methodology? ▾
    The complete CMS methodology is published in: Syed, A. H., Habeebi, S. A., & Habibi, S. M. M. (2026). From Stochastic to Deterministic: A Multi-Criteria Decision Analysis Framework for Bounded Semantic Parsing in AI-Driven Recruitment Screening. Expertini Research Department. Available at: research.expertini.com →

    Academic References

    • [1]Syed, A. H., Habeebi, S. A., & Habibi, S. M. M. (2026). From Stochastic to Deterministic: A MCDA Framework for Bounded Semantic Parsing in AI-Driven Recruitment Screening. Expertini Research. View paper →
    • [2]Belton, V., & Stewart, T. J. (2002). Multiple Criteria Decision Analysis: An Integrated Approach. Kluwer Academic Publishers.
    • [3]Köchling, A., & Wehner, M. C. (2020). Discriminated by an algorithm: A systematic review of discrimination and mitigation strategies in automatic recruitment filtering. Business & Information Systems Engineering, 62(6), 615–628.
    • [4]Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg more employable than Lakisha and Jamal? American Economic Review, 94(4), 991–1013.
    • [5]Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating bias in algorithmic hiring. Proceedings of FAccT 2020 (pp. 469–481).
    • [6]Ji, Y. et al. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38.
    • [7]Robertson, S., & Zaragoza, H. (2009). The probabilistic relevance framework: BM25 and beyond. Information Retrieval, 3(4), 333–389.
    • [8]Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers. Proceedings of NAACL-HLT 2019 (pp. 4171–4186).

    Start scoring candidates the right way

    Deterministic. Bias-free. Fully auditable.
    Powered by Gemini AI semantic extraction and MCDA-bounded mathematics.

    ⚡ Score a Candidate Now → Classic Resume Score Explore ERIS →