Rolewise Resume Matching

Overview & result

Rolewise retrieves job descriptions, sends each resume–job pair to Jev, and asks for one suitability decision; the author reports 838 evaluations in 61.3 seconds with 20 requests in flight.

From the original post

Spent some time testing #Jev from @typesafeai on a use case I care about: matching a resume against real job descriptions.

I built a small demo called Rolewise. @metix_ai handles job retrieval; Jev handles the fit decision.

The setup:

- Parse the resume and generate editable search filters.

- Retrieve jobs and full descriptions through the Metix AI API.

- Send each resume–JD pair to Jev, with 20 requests in flight.

- Ask for one decision: suitable or not suitable.

One run: 838 jobs evaluated in 61.3 seconds. That excludes parsing and retrieval.

I started with multiple scoring dimensions, then simplified the task to a single fit decision. For this demo, I wanted to see whether Jev could help someone narrow a list of opportunities without adding another scoring system to interpret.

The recording shows actual requests completing—no simulated progress. Retrieval and evaluation are separate steps, so you can inspect the jobs before running Jev.

This is a throughput result, not an accuracy benchmark. I haven’t established how reliably those decisions agree with a human reviewer yet. That’s the next thing I’d test, especially on borderline cases.

Sharing this as a concrete Jev example. If you’re building something similar, I’d be curious how you’re defining “fit” and checking the results.

Job data/API: https://platform.metix.ai

How Jev fits in the loop

1. A retrieval service supplies job descriptions and filters.
2. Jev makes one bounded suitability decision for each resume–job pair.
3. The application uses the decisions to narrow the opportunity list and sends uncertain cases to review.

How to reproduce

1. Separate resume parsing and job retrieval from the decision stage.
2. Run a small fixed set of resume–job pairs with one suitable / not-suitable question.
3. Add human labels for borderline cases before using the result for ranking or outreach.

Why this build matters

It is a clear example of simplifying a fuzzy matching workflow into a single typed decision and measuring throughput without pretending that throughput proves accuracy.

Reported performance

Latency: 61.3 seconds for 838 jobs in the cited run

Limitations

• No human-agreement or accuracy result was reported in the reviewed post.
• Resume and job data can contain sensitive personal information and needs appropriate handling.
• The cited throughput should not be generalized to a different retrieval provider or concurrency level.

Patterns