How Our AI Algorithm Actually Works

The decision to use TF-IDF rather than BM25 (Best Match 25) or a pure transformer embedding was deliberate. BM25 improves on raw TF-IDF by introducing a document-length normalization parameter, which matters in information retrieval over long documents. In the resume-to-job-description matching context, however, the asymmetry in document length is predictable and bounded — resumes are typically 400–900 words; job descriptions 200–600 words. BM25's saturation parameter provides marginal benefit over this narrow range. The implementation uses scikit-learn's TfidfVectorizer with a custom stop-word list tuned for HR language ("responsible for," "proven track record," "strong background in"), sub-linear TF scaling enabled, and unigram-plus-bigram tokenization to capture two-word technical terms ("machine learning," "project management," "cross-functional") that single-token analysis would miss.

The corpus used to compute IDF weights is a rolling dataset of job descriptions ingested from publicly available postings. This corpus is updated periodically rather than trained on a static snapshot. The practical effect: when a technology term becomes ubiquitous (say, "AI" or "cloud"), its IDF weight declines because it now appears in most postings, and its discriminative power decreases accordingly. The model adapts to this without manual retuning.

Critical keywords are operationally defined as terms where the job description's TF-IDF weight exceeds a threshold derived from the distribution of weights in that specific document — typically, the top 20–30% by weight. This threshold is document-relative, not fixed, which means "Python" is critical for a software engineering role (high weight in that posting) but merely preferred for a data analyst role where the description also emphasizes "SQL," "Tableau," and "stakeholder communication" with equal weight. The distinction matters for the optimizer's prioritization: the methodology targets 80%+ critical keyword coverage as the primary objective, with preferred keywords addressed secondarily.

One structural limitation worth naming: TF-IDF operates at the surface form level even after lemmatization. It cannot capture semantic similarity between "revenue growth" and "top-line expansion," or between "P&L ownership" and "budget accountability." That is precisely where Layer 2 — GPT-4 content quality evaluation — compensates, by reasoning over semantic equivalence that statistical methods cannot reach. The two layers are complementary by design, not redundant.

The term "ATS score" is widely used as though it refers to a single number produced by a single system. In practice, a corporation using Workday may configure their tenant to weight the most recent position's responsibilities at 60% of the skills match calculation, while another employer using the same Workday platform weights all positions equally. Both configurations are valid within Workday's system, and both produce different outputs for the same resume against the same job description.

Research by Chadda et al. (IEEE Access, 2018) and subsequent work by Bevara et al. (MDPI Electronics, 2025) on transformer-based resume embeddings consistently shows that semantic matching methods outperform pure keyword approaches in cross-platform evaluation — precisely because they are less sensitive to this configuration variance. A resume that communicates genuine skill in a domain does so through multiple linguistic signals, not just exact-match vocabulary. Semantic signals are more robust to configuration differences than keyword counts.

The TalentTuner ATS Match Model addresses configuration variance in two ways. First, the scoring is deliberately calibrated against the midpoint of observed configuration ranges, not against any single employer's settings. Second, the feedback the model provides — specifically the identification of missing critical keywords and weak content areas — is actionable regardless of the target employer's specific configuration. Adding a high-weight keyword improves performance across all plausible configurations; improving achievement framing improves performance wherever GPT-4-style content evaluation exists (and recruiter judgment is a de facto version of that evaluation even on platforms that do not use LLM scoring).

The recency layer (Layer 5) is where configuration variance has the least impact. Across every ATS platform, and in direct human review, a resume that prominently features skills and achievements from the last 2–3 years outperforms one where equivalent skills are buried under 8-year-old experience. This is one of the most consistent signals in the dataset and one of the most under-discussed in conventional ATS optimization guidance.

Three parsing edge cases produce disproportionate damage relative to their frequency among the 50,000+ resumes analyzed.

Skill bar charts and infographic elements. A significant subset of design-forward resumes includes visual "skill bars" — horizontal bars indicating proficiency level (e.g., "Python: 80%"). These are rendered as vector graphics or embedded images. No ATS parser currently reads graphic elements for text content. The skill name is invisible to the ATS even if it appears visually prominent to a human reviewer. Every skill represented only in a chart is a missed keyword match from the ATS's perspective. TalentTuner flags this pattern and counts the skills toward the missing-keyword gap rather than the matched-keyword count.

Multilingual resumes. Candidates who work across language markets sometimes include section titles or skill descriptions in multiple languages. spaCy's language detection pipeline identifies the primary language of the document and flags secondary-language content as a parsing risk. ATS systems without multilingual normalization may fail to tokenize foreign-script content correctly, causing section boundaries to collapse. The practical recommendation: maintain a single-language, single-script document for ATS submission, even if a multilingual version is appropriate for direct recruiter contact.

Scanned PDF documents. A non-trivial fraction of resumes submitted to TalentTuner arrive as scanned images wrapped in a PDF container — typically because the candidate has exported from a legacy word processor, printed, and re-scanned. PyMuPDF's image detection layer identifies these documents before any text extraction is attempted. The system returns a parse-failure warning rather than generating a misleadingly low score from garbled OCR output. Candidates receive a specific recommendation to export from the source document rather than re-scanning. This edge case accounts for a disproportionate share of "I got a 0% score" support contacts, and the detection logic was added specifically to intercept that experience.