An engineer who can build a model isn't automatically one who should be trusted with protected health information. The screen that actually matters is judgment under HIPAA, not ML or Python skill level, and most job descriptions never test for it.
See the screening checklistMost healthcare AI job descriptions and interview loops are built to test the same things any ML or backend role tests: model quality, system design, code craft. Those matter, but they don't tell you whether a candidate can be trusted with protected health information (PHI). That's a separate skill, and it's rarely screened for on purpose.
HIPAA itself doesn't hand employers a technical skills test. It expects access to PHI to be limited to the right people based on role and risk, and it leaves the actual judgment call to the hiring company. In practice, that means the burden falls on the interview process to surface whether a candidate's instincts around access, de-identification, and ambiguous requests are sound, before they're ever in a position to make a costly mistake.
Five areas separate a candidate who's genuinely ready for regulated healthcare data from one who just has the ML or backend skills to look ready on paper.
| What to screen for | How to actually test it | The red flag |
|---|---|---|
| De-identification method knowledge | Ask them to explain the difference between Safe Harbor (removing the 18 HIPAA-specified identifiers) and Expert Determination (a qualified expert formally certifying re-identification risk is very small) | Treats "we stripped a few obvious columns" as a complete answer |
| Access discipline | Ask how they'd design a system so most of the team never needs direct PHI access at all | Can't explain why broad admin access is a liability rather than a convenience |
| Audit and logging habits | Ask about a system they've built or operated where every PHI access was logged and reviewable after the fact | Treats audit logging as something to bolt on later, not design in from the start |
| BAA awareness | Ask what a Business Associate Agreement actually obligates a vendor, contractor, or sub-processor to do | Has never encountered the term in a role that touched regulated data |
| Judgment under ambiguity | Scenario: an unfamiliar internal contact asks for "a sample dataset for a demo." What do they do first? | Prioritizes shipping speed over verifying the request and its scope |
This screen only works if the candidate pool already contains people who've touched regulated data before. A general AI/ML sourcing pass built around model architecture keywords rarely surfaces them, because this experience doesn't show up as a flashy GitHub repo, it shows up as a quiet line on a resume about working inside a HIPAA-covered pipeline.
The engineers who are actually good at this don't lead with it. It shows up as one line about de-identification pipelines or claims data, buried under more conventional ML or backend work. You have to know to go looking for it.
Sourced entirely through outbound for a de-identification pipeline covering more than 10 million patients' worth of data: 633 candidates, 35 days, closed at $140K total comp.
A narrow requirement: healthcare data experience (EHR, claims, registries) plus SQL and Python fluency. 91 candidates, our fastest close at 22 days, $185K total comp.
Judgment under HIPAA specifically, not general ML or Python skill. That means testing whether the candidate understands de-identification correctly (Safe Harbor versus Expert Determination) and defaults to least-privilege access on their own, without being told to.
Not strictly. Direct PHI experience is a strong signal, but the real test is whether a candidate asks the right questions when they haven't seen a specific regulation before, not whether they've memorized HIPAA's text.
Describe an ambiguous access request, such as an unfamiliar internal contact asking for "a sample dataset for a demo," and see whether the candidate's first instinct is to verify the request and its scope or to just ship it.
A general ML screen tests modeling and engineering skill. A regulated-healthcare screen adds a second, separate axis entirely: whether the candidate can be trusted with protected health information, which depends on access discipline and judgment, not model accuracy.
Based on current market postings for hands-on privacy and de-identification engineering roles, roughly $125,000 to $150,000 base is a reasonable planning range for an individual contributor. It moves higher with seniority or if the role layers ML modeling responsibilities on top of the privacy engineering work.
David has run healthcare-AI and regulated-data searches directly, including both Dandelion Health placements referenced above, which is why this checklist comes from what the screen actually needs to catch, not a generic compliance template.
Book a 15-minute call. We'll tell you honestly what the candidate pool actually looks like for this exact hire.
Book a call