Sentari
Methodology

How we score hiring risk.

Sentari produces probabilistic risk indicators — never accusations. This page documents what we measure, how the model behaves, and where the limits are.

Inputs
URL · Text · Email
Model class
LLM + heuristics
Output
Score 0–100 + signals

What we analyze

  • Payment requests

    Mentions of fees, deposits, equipment costs, or training charges borne by the applicant.

  • Off-platform pressure

    Push to WhatsApp, Telegram, Signal, or personal email — particularly early in contact.

  • Company verifiability

    Domain age, public footprint, executive presence, and consistency with claimed scale.

  • Recruiter authenticity

    Identity signals, employment history coherence, and historical posting patterns.

  • Posting linguistics

    AI-generated, copy-pasted, or template patterns common to high-volume scam networks.

  • Compensation outliers

    Salary, equity, or benefits significantly above market for the stated role and seniority.

  • Process irregularities

    Unusually fast offers, missing interview steps, or pressure to commit within hours.

  • Data harvesting cues

    Requests for SSN, banking info, or government IDs prior to a verified offer.

How a score is produced

  1. 1
    Normalize input
    URLs are fetched server-side with a 5s timeout and 1MB cap. Text input is sanitized. Personal data in pasted content is not persisted beyond your scan record.
  2. 2
    Extract signals
    The model classifies content against the indicator set above, returning structured evidence — not free-text conclusions.
  3. 3
    Aggregate score
    Indicators are weighted into a 0–100 composite. Bands: Low (0–34), Elevated (35–64), High (65–100).
  4. 4
    Return recommendations
    Specific, actionable next steps — verify the recruiter, request a counter-signal, decline the off-platform move.

Limitations

  • Probabilistic, not judicial. Scores estimate risk patterns. They are not findings of fraud and must not be used as the sole basis for any employment decision.
  • Coverage gaps. Closed platforms (LinkedIn private postings, internal ATSes) can't be fetched server-side. Paste the content instead.
  • Adversarial drift. Scam patterns evolve. We retrain regularly and rely on community reports to surface new patterns.
  • False positives exist. Legitimate roles can trip indicators (high comp, fast process). Always weigh evidence in context.

Saw a pattern we should learn from?

Submit a report. Confirmed patterns retrain the model.

Report a job