Methodology
Every ScorAxis score is the sum of two halves: a deterministic rubric and a small set of AI checks. The deterministic half is the ground truth — same input, same output, every time.
Deterministic rubric (v1, frozen)
Six checks. Each runs on the parsed JSON, never on the raw PDF, so results don't drift across re-uploads of the same content. The rubric is versioned and frozen for reproducibility — the same resume produces the same score until we publish a v2 and bump the version number alongside it.
- Contact information. Name, email, and phone are present and parseable.
- Sections present. Experience, education, and skills are all detected as distinct sections.
- Word count band. The resume falls inside the 350–900 word range that ATS parsers handle best.
- Bullet-point structure. Experience entries use bullet points rather than free-form paragraphs.
- Date-format consistency. Date ranges follow a single, machine-parseable format throughout.
- Parsing-friendly layout. The PDF is single-column, no embedded tables, no images masquerading as text.
| Check | Max points |
|---|---|
| Contact information | 10 |
| Sections present | 15 |
| Word count (350–900) | 10 |
| Bullet-point structure | 10 |
| Date-format consistency | 10 |
| Parsing-friendly layout | 15 |
| Total deterministic | 70 |
AI checks (5 dimensions)
On top of the rubric, an LLM evaluates five qualitative dimensions that recruiters actually notice. The model has access only to the parsed text — never the raw PDF, never anything else from your account.
- Action-verb strength. Strong verbs (“led”, “shipped”, “migrated”) instead of weak ones (“helped with”, “was responsible for”).
- Quantified impact. Numbers, percentages, and named outcomes — not vague generalities.
- Keyword density. Coverage of the role-specific terms an ATS keyword filter scans for.
- Tense consistency. Past tense for previous roles, present tense for the current one; consistent across each entry.
- Buzzword overuse. Generic filler (“synergy”, “rockstar”, “dynamic”) flagged so you can replace it with something concrete.
Each AI issue is assigned a severity: high (−4 pts), medium (−2 pts), or low (−1 pt). The AI score starts at 30 and deductions are applied until it reaches 0. Up to ten issues can be flagged per scan.
AI checks may be temporarily unavailable while we expand capacity. When they are, your deterministic score is unaffected and the issues list filters out AI-only findings until the model is back online.
Where your data lives
Scores never leave our infrastructure. Resume PDFs are stored in S3 with server-side encryption (SSE-S3). Parsed JSON sits in a private RDS Postgres database in ap-south-1. Presigned upload URLs expire in five minutes.