Methodology
How the recruiter check works
Applye reads a job the way a screener does, not the way you hope it reads. The check is built in layers: cheap deterministic code first, then rubric-guided AI judgement only where it earns its tokens. This is the career-ops methodology, rebuilt for a desktop GUI.
The score
Every scored job gets a single fit reading on a 0 to 100 scale, plus a one-line verdict you can act on: pursue it, hold and improve the application, or skip it. The number is not the point; the point is a clear, blunt signal and the reasons behind it. The line between "pursue" and "hold" is defined by the scoring rubric and shown per job, not a fixed universal cutoff.
The tone is deliberately direct: how an HR screener actually reads, not encouragement.
Layer 1: deterministic, 0 tokens
Before any model runs, plain code does the obvious work for free. It parses the posting, hard filters on location, salary, contract type, and visa, and assigns a legitimacy tier. Obvious mismatches and likely-fake postings stop here, so you never spend tokens scoring a job that was never real.
Green
Looks legitimate. Clear company, comp, and posting signals. Proceeds to scoring.
Yellow
Flagged. Missing salary, thin detail, or odd patterns. Scored, but with a caution.
Red
Ghost-job patterns or a suspicious domain. Surfaced as a warning, not quietly scored.
Scan summary, for a batch: "2 passed, 1 rejected (location), 1 flagged yellow (no salary)."
AI scoring layers
Only jobs that pass the code filter reach the model, and it is asked to act as a blunt recruiter, not a cheerleader. Each result is structured and cached against a hash of the job text, so re-opening it costs nothing.
- Recruiter rubric score with point-by-point deductions, not a vibe.
- Top missing keywords the posting expects and your profile lacks.
- Hiring-manager red flags a screener would raise.
- ATS filter pass on the formatting that silently breaks parsers (the real hyperref lesson, encoded as a check).
Before you submit
The check closes with practical notes, so you fix the application before it goes out:
- Salary missing prompts a comp-research step.
- Portfolio required is called out.
- Deadlines are surfaced.
Honest about judgement
This is rubric-guided LLM judgement, not a deterministic formula. Two runs can phrase things differently, and the model is an excellent assistant but a poor judge of someone's life. So the score is a sharp opinion you can argue with, never a verdict. The deductions are shown, you can expand the full breakdown on demand, and you always override it. We do not publish fixed rubric weights as if they were physical constants; the rubric lives in auditable skill files that ship with the app and open to the public with the source.
Token economy
The whole design follows one rule: do not call AI where code suffices. Parsing, filtering, legitimacy tiers, the first-pass ATS check, follow-up badges, and analytics are all plain code at 0 tokens. The model is reserved for genuine judgement (scoring, tailoring), runs on a cheap tier for routine work, and every answer is cached. A real search costs cents, not a subscription.
The augmentation boundary
The check scores and the app drafts, but it stops there. It never decides for you and never submits for you. A recruiter is a person, and the relationship is yours, not a bot's. Drafting is automated. Submitting is not.