Healthcare staffing
Every recruiter note, sorted in 86 ms.
Booking lane, credential status and unit pool. Taught by your recruiters.
One note. Three typed answers.
Synthetic note. Real Jev outputs.
Recruiter note
Jules, RN, 9 yrs, Boise, is availble for the medical ICU opening: 7 weeks, coming from out of town, stipend offered. TB is 15 months old and NRP lapsed last week. Rate aside: asking $68/hr; references ready.
| Question | Before feedback | After feedback |
|---|---|---|
assignment_ | travel_contract 0.99 | per_diem 0.93 |
credential_ | verify_first 0.28 | do_not_submit 0.99 |
unit_ | critical_care 1.00 | critical_care 1.00 |
Agency rules: under 8 weeks is per diem. Lapsed certs stop submission.
20 reviewed notes. 28 new ones.
Recruiters corrected 15 answers. Rotor did the rest.
| Held-out notes | Before | After |
|---|---|---|
| All answers | 50% | 82% |
| Booking lane | 32% | 68% |
| Credential status | 39% | 100% |
| Unit pool | 79% | 79% |
| Notes fully correct | 4 of 28 | 15 of 28 |
29 answers fixed. 2 regressed. Policy compile cost about $0.05.
Same accuracy. A fraction of the wait.
Same notes, same feedback, general reasoning LLM.
82% vs 83%Accuracy with feedback
86 ms vs 1,682 msMedian response time
$0.07 vs $3.34Model cost per 1,000 notes
Jev versus gpt-6-sol, low reasoning. List prices. Rotor fees excluded.
No prompt-tuning sprints.
Recruiters correct answers. Engineers approve results.
Hand-built LLM loop
- Export misses from logs
- Label them by hand
- Rewrite the prompt
- Rerun the eval set
- Hunt for regressions
- Redeploy
Rotor loop
- Recruiters confirm or correct answers
- Rotor writes a per-question policy
- Rotor compares versions on saved tests
- You approve the new version
Estimate your savings.
An estimate. Change the assumptions.
- Model cost per month
- $3.56 vs $167
- Waiting removed per month
- 22 hours
- Engineer hours returned per month
- 15
- Engineering cost returned per month
- $1,800
Assumes 30 minutes of engineer review per Rotor iteration.
What we measured, and what we didn’t.
- Synthetic notes. Test on your own notes first.
- Feedback teaches only rules recruiters actually correct.
- “Exactly 8 weeks” regressed twice after feedback.
- One run per note. Jev varies slightly between runs.
- Model weights never change. Corrections become context.
Bring 20 reviewed notes.
We’ll measure your rules on your data.
Build with us