jobbyjev

Give it a resume and a list of companies. It ranks the companies by how likely the candidate is to get an interview, puts a confidence on each, and names the mismatch when there is one. A local reproduction of Backdoor's "picking the best company using Jev" demo, built on TypeSafe's Jev model.

python3 rank.py samples/sample_resume.md          # 400 companies, ~25 s, ~\$0.03
python3 rank.py my_resume.pdf --limit 50          # a quick look
open out/report.html

Which way round it runs

It ranks companies, for one candidate. You give it your own resume and it tells you where your applications are worth spending. That is the direction it was built for and the only one anyone has looked at it in.

Do not turn it around and screen candidates with it. Run many resumes past one company and would_interview becomes a hiring filter, and it would be a bad one. Three reasons, all measured rather than asserted:

• It is generous. The median would_interview across 400 companies was 0.80, and 88 cleared the 0.70 "likely" line. The demo this reproduces found one in 400. Something that says yes to a fifth of the world sorts nobody.
• Nobody has checked a single one of these probabilities against who actually got an interview. They are Jev's numbers about a resume and a company description, not evidence about a person.
• The companies it scores against are made up in the ways that matter. See the data note below.

A number nobody has validated, applied to somebody's job application, is worse than no number, because it looks like it knows something.

Needs Python 3.12, requests, and a TypeSafe key in TYPESAFE_API_KEY (or .env, see .env.example). PDF resumes need pdftotext (poppler).

How it works

One Jev request per company. The state is {candidate: {resume}, company: {name, hq, stage, size, category, description, hiring_for, founder_led}} and every request carries the same six judgments (questions.py):

id	type	judgment
domain_fit	Score 0..3	does the candidate's past work match what the company builds
role_fit	Score 0..3	does their discipline match a role in hiring_for
stage_fit	Score 0..3	have they worked at employers of this stage and size
would_interview	Noul	would a recruiter here reach out
location_ok	Noul	is the stated location or preference compatible with the HQ
mismatch	Choice	the single biggest disqualifier, with an explicit none

Code combines them (questions.compose): fit is a weighted mean of the three Scores, chance is half would_interview and half fit, scaled by location_ok. Confidence is the mean of the Score and Choice confidences. A mismatch is only reported when the Choice confidence clears 0.5. The weights and thresholds are constants at the top of questions.py; change them and re-run, the response cache makes that free.

Why one request per company and not 40 companies per request: the jev-orderby-bench measurement found the same rows through a 40-rows-per-state layout fail the ranking gate (pairwise inversion 0.171 against a 0.15 threshold) while one row per request passes. Ranking is the whole product here.

Output

out/results.json holds the summary and every company's raw signals. out/report.html is a self-contained page: the company grid tinted by chance with a red ring on every company over 50%, the signal meters for the selected company, a sideways-scrolling shortlist of everything over the line, and a full ranked table. Pass --photo headshot.jpg to put a picture in the header; it is inlined, never copied into the repo. Logos come from Google's favicon endpoint, so the page needs network for logos only.

A first run

Sample resume (fictional payments backend engineer, Taipei) against all 400 companies, jev-1.13.0, 16 workers:

checked	likely interview	mismatch flagged	avg chance	time	input tokens	cost
400	88	90	0.55	22.7 s	669k	$0.028

Top of the list: Modern Treasury, Column, Unit, Melio, Increase, XREX, Mercury, Supabase. Bottom: the semiconductor companies, all flagged as a discipline mismatch. That is the order a person would give.

What it is not

• Generous on "would interview". The median would_interview probability across the 400 was 0.80, so half the companies come out as plausible and 88 clear the 0.70 "likely" line. Backdoor's demo shows one "would hire" out of 400. Either raise LIKELY_THRESHOLD, or add the signal this build lacks: live openings. Without a posting to match against, "would a recruiter reach out" is answered about the company in general, and most companies hire backend engineers in general.

• Not calibrated on hiring outcomes. The probabilities are Jev's, and nobody has checked them against who actually got interviews. The same bench above found Jev underconfident on easy labels and failing four of six gates on a hard graded-relevance probe. Treat the order as a shortlist generator, not a prediction.

• Nothing in the company data was researched. data/companies.json is 400 well-known tech companies written from a language model's general knowledge on 2026-09-17. No entry was sourced or checked against any company's own material. stage, size, category, hiring_for and founder_led are impressions rather than facts, and some of them are certainly wrong. Do not quote an entry as a fact about the company it names, and do not read anything into which companies are present. The file says all of this in its own _provenance record, because a JSON file gets copied and loaded by things that never read a README. It also has no job postings, so the demo's "they're hiring now" signal has no equivalent here. Replace it with sourced data before any answer matters.

• English only, effectively. Jev's accuracy on Chinese-language resumes is documented as lower. Translate first.

• Resume text is sent to TypeSafe. Read their data terms before using a real resume.

Files

• rank.py CLI and pipeline
• questions.py the six judgments and the composite policy
• jev_client.py HTTP client with retries and a SQLite response cache
• report.py HTML renderer
• data/companies.json the 400 companies
• samples/sample_resume.md a fictional resume so it runs out of the box
• test_questions.py offline tests, python3 test_questions.py

The rest of the series

All mine, so none of them is independent corroboration of the others.

• jev-orderby-bench measures whether Jev's probabilities can be sorted on at all. It decided this project's architecture: the same rows batched 40 to a request failed its ranking gate at 0.171 inversion against a 0.15 threshold, where one row per request passed. Ranking is the whole product here, so jobbyjev sends one request per company and pays for it.
• commitjev reviews commits, and ships labelled fixtures that say what it catches.
• jevq checks Jev questions against the model's documented failure modes without calling the API. Run it here with jevq --from questions:QUESTIONS. The first time anyone did, it produced three findings, and all three were wrong: it read "at least one role" as a counting question when it is really an existence check. Fixing that was worth more to jevq than the run was to this repo.

MIT licensed.