Technical story

How the scoring system works

job-eval.com is powered by a specialized ModernBERT-based scoring model built for resume-to-job-description evaluation. It is not a generic prompt wrapper around an expensive frontier model.

1. A specialized ModernBERT scorer

The production evaluator uses a fine-tuned ModernBERT encoder and a multi-head scorer trained to predict four explicit 1-to-5 job-fit signals rather than one opaque output.

2. Four decision heads

The model scores requirement coverage, seniority fit, domain alignment, and ATS-style keyword overlap separately. The product then surfaces the average while keeping the sub-scores visible.

3. Explainability added at the model layer

Instead of relying on another model to “explain” the result after the fact, the system uses the trained pooler attention from the scorer itself and aggregates it into readable paragraph-level evidence.

What makes this different

Why not just send every resume and job description to a frontier model?

Frontier models are powerful, but they are expensive, less predictable in repeated screening workflows, and harder to operationalize for a narrow scoring problem. job-eval.com takes the opposite approach: train a dedicated model for a dedicated task, then expose the evidence behind the score.

The shipped production scorer is the phase 5 checkpoint. It keeps the phase 4 architecture, but improves fit quality while widening the attention maps so the explainability view is actually readable in a product UI.

Explainability uses span-level attentionAttention is aggregated over paragraphs instead of individual tokens so users can read the evidence directly.

Entmax 1.5 improves readable highlightsThe phase 5 configuration uses entmax-style sparse attention so irrelevant tokens can drop out instead of all tokens receiving non-zero weight.

Forward-pass explanation pathThe deployed system uses the scorer’s own attention maps rather than adding extra backward-pass explainers that raise latency without clear benefit.

Build credibility into the product

Technical enough for experts. Simple enough for candidates.

The product surface is intentionally clear, but the underlying system is grounded in a trained ModernBERT architecture, explicit scoring heads, and a documented explainability method.

Try the evaluator Meet the founder