RigelLog in

AI Transparency

Last updated: April 16, 2026

Rigel is built on a core promise: every matching score should be traceable to specific, inspectable inputs. This page explains, in plain language, where and how we use AI in the Service — and, just as importantly, where we don't.

TL;DR

  • AI is used for CV parsing only. We use large language models to extract structured information (experience, education, skills) from unstructured CV documents.
  • AI is NOT used for scoring. Match scores are computed by deterministic algorithms with explicit, inspectable rules and weights. No LLM decides whether you are a good fit.
  • Every score is reproducible. Run the same candidate against the same job twice, you get the same number. Every time.
  • You can see the math. Both candidates and recruiters can open any score and see which skills matched, which didn't, and how each section contributed.

1. Where we use AI

1.1 CV parsing

When you upload a CV, we need to extract structured information from what is often free-form text or complex formatting. We run a multi-stage pipeline that combines deterministic rules (regex, layout analysis) with large language model calls for ambiguous extraction tasks — distinguishing role titles from company names, normalizing skill variants, parsing date ranges, and similar tasks.

We currently use Google Gemini 2.5 Flash, hosted on Google Cloud Vertex AI. Your CV data is processed within Google Cloud and is not used to train Google's models (per Google Cloud's terms for Vertex AI).

You review the extraction before it is used. After parsing, you are shown a full draft profile and can edit every field before confirming. Nothing enters the matching pool until you approve it.

1.2 Job description generation (recruiter-side, optional)

Recruiters can optionally ask Rigel to generate a polished job description from the structured job data they have entered (skills, weights, experience requirements). This is a convenience feature — the recruiter reviews and edits before publishing. The structured data, not the generated description, drives matching.

1.3 Skill normalization

When you or a recruiter types a skill, we map it to a canonical entry in our skill ontology (e.g., "Py" → "Python", "React.js" → "React"). This uses deterministic matching first; LLM assistance is used only for genuinely ambiguous cases, and matches are logged.

2. Where we do NOT use AI

2.1 Match scoring

The core matching engine is fully deterministic. A candidate's score for a job is computed by:

  • Skills score: structured comparison of the candidate's confirmed skills against the job's required and preferred skills, with level matching (beginner / intermediate / advanced / expert).
  • Experience score: years of relevant experience against the job's stated requirement, with explicit penalty and bonus rules.
  • Education score: degree level and field match against the job's education requirements.

Each section has a numeric score from 0 to 100. The final overall score is a weighted average using the weights the recruiter set when posting the job (e.g., Skills 50%, Experience 35%, Education 15%).

No LLM or embedding model is in this pipeline. No "semantic similarity" magic. No opaque re-ranking. Same inputs always yield the same score.

2.2 Ranking and filtering

Candidates are ranked strictly by overall score, descending. We do not apply personalization, engagement signals, or any non-score factor to the ranked list. Recruiters see the same order every time, subject only to changes in the underlying data (new applicants, edited profiles, edited job requirements).

2.3 Bias and "fit"

Rigel does not evaluate "culture fit," personality, or any attribute not in the structured requirement list. The scoring engine has no access to candidate names, photos, demographic information, or inferred characteristics. Recruiters see anonymized candidates by design.

3. Data used by AI systems

  • Inputs: the CV you upload, and (for recruiters) the structured job data you enter.
  • Outputs: structured JSON representing your extracted profile (or generated JD text).
  • Not used for training: your data is not used to train third-party AI models. We do not sell, share, or otherwise disclose your data to train external models.
  • Logging: LLM calls are logged with their inputs and outputs for debugging and quality monitoring. Logs are retained for 90 days, access-controlled, and automatically deleted after that.

4. Why we built it this way

Most "AI recruiting" tools use opaque scoring: a large model reads a CV and a job description and emits a number. This has two problems.

First, it's not reproducible. The same candidate can get different scores on different days, and no one — not the recruiter, not the candidate, not the vendor — can explain why.

Second, it's not auditable. If a candidate asks "why wasn't I selected?", or a regulator asks "how does this avoid discrimination?", "the model said so" is not a satisfying answer. And increasingly, not a legal one.

Rigel's architecture — AI for extraction, rules for scoring — gives us both the lift of modern language models (no manual profile entry) and the accountability of traditional software (every score is a sum of numbers you can look at).

5. Your rights regarding automated decisions

Rigel surfaces matches but does not make hiring decisions. Recruiters make all hiring decisions and are responsible for the lawfulness of those decisions under applicable law, including anti-discrimination, equal-employment, and automated-decision-making regulations (such as NYC Local Law 144, EU AI Act, and similar regimes).

You can:

  • See how any match score was computed — open the score drawer for a section-by-section breakdown with the specific skills and requirements that matched.
  • Edit your profile to correct any extraction error before confirming it.
  • Opt out of the matching pool at any time from Settings.
  • Request a full export of your data.
  • Contact us with concerns about any scoring outcome at support@rigelsignal.com.

6. Changes to this statement

If we change how AI is used in the Service — add a new use, switch models, or change data practices — we will update this page and notify active users by email.

Questions? Email privacy@rigelsignal.com.