Resume screening with Jev: match, seniority and next step

What this check scores

The example above screens one candidate for a Senior Backend Engineer role that asks for Go, Postgres and Kubernetes. The candidate has six years of backend experience, mostly in Python and Django, led a migration to Kubernetes at a fintech startup, has built internal tooling in Go for the last 18 months and is comfortable tuning Postgres. Jev answers three questions in one call:

This is a deliberately mixed profile. The candidate’s main language is not the one the role names, but they have recent Go experience and have done the Kubernetes and Postgres work the role needs. That is the kind of case where a keyword filter rejects a good candidate. What to look at is how match spreads across Partial and Strong, and whether next_step leans to phone_screen rather than reject.

The state is short and there are three questions, so this is a standard evaluation: 1 credit on JevStation.

How to write screening questions that work

State the bar, do not imply it. The preset asks "Does the candidate meet the seniority bar?" without saying what the bar is. Jev reads instructions literally, so it will guess. Write the bar in terms of scope and responsibility — for example, has led a system or a migration end to end — in the question’s criteria.

Compute years and dates in code. "Six years" and "the last 18 months" look easy, but TypeSafe documents counting and date comparison as weaknesses. If the role needs a minimum number of years, calculate it from structured data in your applicant tracking system and pass a field such as years_backend: 6, or leave it out of the automated step.

Send only what the decision should use. Jev suffers from context rot: unrelated material in the state costs accuracy. Trim the resume to experience and skills, and leave out anything a hiring decision must not depend on (see below).

Keep reject from being a default. Give each next step a clear definition, as the preset does, and treat reject as a suggestion for a person to confirm, never an action.

For more on phrasing, see Designing Questions Jev Can Answer.

From one candidate to a screening step

1. Start with past decisions. Run 50 to 200 candidates your team has already assessed. The batch page runs one question set over every row of a CSV, so you can compare Jev’s answers with your recruiters’ decisions before anything touches a live pipeline.
2. Use Jev to order the queue. Sort applicants by match and fast-track confident onsite or phone_screen suggestions to a recruiter. Everything else still gets a human read.
3. Route every suggested rejection to a person. A rejection is the decision with the highest cost to the candidate and the least visibility to you. Do not let it fire automatically.
4. Log the answers. Store the probabilities, the question set and the model field every response returns, so you can reconstruct how a candidate was scored. Re-check thresholds when the model or the questions change.
5. Call it from your pipeline. The same evaluation runs through the System One API with an API key and charges the same credits as the playground.

Use it responsibly

Automated screening of people is high-stakes, and a fast, cheap classifier makes it easy to apply one mistake to thousands of applicants. Before Jev scores real candidates:

• Have a person review every rejection. Jev’s output should change the order in which people are reviewed, not replace the review.
• Test for disparate outcomes on your own data. Compare how often candidates from different groups receive each next_step, using data you are permitted to hold, and keep that audit separate from the state Jev sees. A gap you did not expect means the questions or the inputs need work before launch.
• Do not send protected attributes or obvious proxies. Names, photos, ages, graduation years, addresses, nationality, marital or family status and similar details do not belong in the state.
• Remember that calibration is a group property. TypeSafe notes that calibration is measured across groups of predictions and does not guarantee that any individual answer is correct.
• Check local law and your own policies. Rules on automated decisions in hiring differ by place and can change. Confirm what applies to you before an automated step influences who moves forward.

Where Jev is the wrong tool

• You must explain each decision to the candidate. Jev returns probabilities, not reasons. One published analysis of Jev raises exactly this objection for regulated fields.
• The role is decided by hard numbers. Years, dates, certifications with expiry dates and salary bands belong in code.
• You want it to make the final call. It should not. Use it to spend reviewer time where it matters most.
• Most resumes are not in English. Measure that group on its own first.

For a broader comparison with general LLMs and trained classifiers, see Jev vs LLM classification. For scoring the other side of the funnel, see lead qualification.