The Jev community app roundup is blowing up again 🔥

kaucap08’s 12,000 résumé-matching experiment, bit-cook’s X spam filter, ondata’s VisiData judgment plugin, TheAIMogul’s Raspberry Pi JevCam, hasanzaman22’s effort router… Players across the community (all third-party community projects) have turned TypeSafe’s Jev decision model into a next-gen “AI headhunter + comment section street sweeper + spreadsheet judge + talking surveillance camera + cost-saving scheduler”:

1️⃣ kaucap08/jev-job-matching (https://github.com/kaucap08/jev-job-matching…)

Independent experiment: 400 real job postings × 30 fictional résumés, 12,000 evaluations total, in about 197 seconds at roughly $0.17 — of 100 postings manually cross-checked, the top 3 hit 96 of them (author’s own test). “12,000 matches for under two bucks!”

2️⃣ bit-cook/Jev-X-Blocker (https://github.com/bit-cook/Jev-X-Blocker…)

A local-first Chrome extension that uses the Jev decision model for semantic filtering of X posts and comment threads. Rules are written in plain English — it understands slang and emoji code words. Hidden content can be restored on the spot at any time. “No more fishing through comment sections by hand!”

3️⃣ ondata/vd-jev (https://github.com/ondata/vd-jev)

VisiData plugin: asks closed-ended questions about each cell in a text column, then writes back yes/no probabilities, fixed labels, or ordinal scores along with confidence as new columns. Jev is the default, but you can swap in other decision models from OpenRouter. “In terminal spreadsheets, every cell gets its own judge!”

4️⃣ TheAIMogul/jevcam (https://github.com/TheAIMogul/jevcam…)

Raspberry Pi 4 + infrared camera. Describe what you want it to look for in plain English, and it asks Jev once per second what’s in the frame. When a target comes into view, it fires off a “radio DJ-style” blaring alarm. Also ships with a worksite PPE safety monitoring app. “One sentence is all it takes to teach a camera to watch for someone!”

5️⃣ hasanzaman22/jev-effort-router (https://github.com/hasanzaman22/jev-effort-router…)

Claude Code plugin: before running each prompt, it first asks Jev how much thinking is needed and how risky the change is, then runs with the lowest sufficient effort. Simple questions go low, debugging and high-risk changes go high or xhigh — no model swap needed. “Not a single extra token wasted!”

kaucap08/jev-job-matching

Source: https://github.com/kaucap08/jev-job-matching

Jev job-matching experiment

An independent experiment: can the Jev model (by TypeSafe AI) match real job postings to the right candidates?

I took 400 real job postings from a public Kaggle dataset and asked Jev to score each one against 30 fictional résumés, for a total of 12,000 candidate–job evaluations.

Disclaimer: I have no affiliation with TypeSafe AI. I paid for all API calls with my own credits. All 30 candidates are fictional. Any resemblance to real people is coincidental.

Results

The top-3 metric was computed on 100 postings with a hand-checked answer key (see below):

• 86 postings have a correct candidate. In 82 of them, that candidate was in Jev’s top 3.
• 14 postings have no suitable candidate (answer “nenhum” = none). Jev’s best score stayed below the 1.5 cutoff in all 14, so it correctly said “none”. These count as hits.

Other numbers from the same check:

You can recompute all of these without calling the API:

``bash python avaliar.py

### About the answer key

`gabarito_claude.csv` was **labeled by Claude** (Anthropic's AI assistant) using the résumés and trap notes, and then **spot-checked by me on a sample**. It is not a fully human-labeled answer key, so read the numbers with that in mind.

### What went wrong

- **Seniority.** Jev often matched the right *field* but the wrong *level*. Without the cutoff there are 26 top-1 misses, and **16 of them** are seniority mix-ups:
  - In 5 "Assistant Store Manager" postings (plus 1 sales/customer-service manager posting), the Store Manager (David) ranked above the Assistant Store Manager (Jasmine).
  - In 3 entry-level retail/cashier postings, the Assistant Store Manager (Jasmine) ranked above the Retail Sales Associate (Emily).
  - For "Assistant Controller", the Controller (Linda) ranked above the Senior Accountant (Raj).
  - It also went the other way: for an "Endoscopy Registered Nurse" posting, the new-grad RN (Samantha) ranked above the experienced ICU nurse (Grace).

  The correct candidate was usually still in the top 3, which is why top 3 (96/100) is much better than top 1 (45/100).
- **The 1.5 cutoff was too high.** Fit scores range from 0 to 3, and `avaliar.py` treats a best score below 1.5 as "no good candidate". Jev's scores were lower than expected: the median best score across all 400 postings was **1.31**, and **225 of 400** postings fell below 1.5. In the answer key, **39 of the 86** postings that *do* have a correct candidate fell below the cutoff and were turned into "none". That is the main reason top 1 drops from 60/86 (no cutoff) to 45/100.

## How it works

| Step | File | What it does | Calls the API? |
|---|---|---|---|
| 1 | `amostra.py` | Samples 400 postings from `postings.csv`: descriptions ≥ 500 characters, 30 per area across 12 areas, then a random fill (`random_state=42`). | No |
| 2 | `curriculos_ficticios.md` | 30 fictional résumés. Some are deliberate "traps": overqualified, keyword-only match, location-locked, part-time only, missing license. The expected behavior is in the "Trap notes" section at the end, which is **not** sent to the model. | — |
| 3 | `rodar_jev.py` | For each posting, asks Jev for a 0–3 fit score per candidate plus five yes/no checks (full-time, remote, Denver, Miami, RN license required). The yes/no checks remove candidates whose constraints the job breaks, then the top 3 are kept. | **Yes** |
| 4 | `gabarito.py` | Picks 100 of the 400 postings for the answer key (writes an empty template). | No |
| 5 | `avaliar.py` | Compares `resultados.csv` with `gabarito_claude.csv`. | No |

Notes:
- The "marketing" area has 29 postings instead of 30. One sampled marketing posting had already been picked for another area and was removed as a duplicate, so the random fill has 41 postings.
- `exploracao/` holds early scripts: a look at description lengths and a 2-candidate smoke test (`teste_jev.py` calls the API).

## Reproduce

Tested with Python 3.14. You need a TypeSafe API key only to re-run `rodar_jev.py`.

``bash
python -m venv .venv
.venv\Scripts\activate            # Windows (Linux/macOS: source .venv/bin/activate)
pip install -r requirements.txt

Check the published results (free):

``bash python avaliar.py

**Re-run the whole experiment (paid, ~US$ 0.17):**

1. Copy `.env.example` to `.env` and put your key in it.
2. Download `postings.csv` from the LinkedIn Job Postings dataset (https://www.kaggle.com/datasets/arshkon/linkedin-job-postings) into this folder. I downloaded it in October 2026. Kaggle updates this dataset, so a different version will produce a different sample.
3. The scripts refuse to overwrite existing outputs, to protect the official data. Move `vagas_400.csv` / `resultados.csv` elsewhere first if you want to regenerate them.
4. Run `python amostra.py`, then `python rodar_jev.py`.

Model outputs may differ between runs and model versions (`jev-latest`).

## Files

| File | Description | License |
|---|---|---|
| `vagas_400.csv` | The 400 sampled postings | CC BY-SA 4.0 |
| `resultados.csv` | Jev's scores, yes/no checks and top 3 for every posting | CC BY-SA 4.0 |
| `gabarito_claude.csv` | 100 postings with the expected best candidate (or "nenhum") | CC BY-SA 4.0 |
| `curriculos_ficticios.md` | The 30 fictional résumés and trap notes | CC BY-SA 4.0 |
| `*.py` | Code | MIT |

## Credits and license

- **Dataset:** LinkedIn Job Postings (2023 - 2024) (https://www.kaggle.com/datasets/arshkon/linkedin-job-postings) by **arshkon** on Kaggle, licensed under CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/). The original `postings.csv` is not included (~517 MB).
- **Code:** MIT License (see [`LICENSE`](LICENSE)).
- **Data:** the CSV files are derived from the dataset above (sampled, with columns added). They are distributed under **CC BY-SA 4.0**, together with the fictional résumés. See [`LICENSE-DATA`](LICENSE-DATA) for details and the list of changes.