First, a disambiguation: the name is crowded. This is Heidi Health (heidihealth.com), the Melbourne-founded clinical AI startup that drafts a clinical note from a doctor-patient conversation, not the children’s novel or any consumer brand that shares the name. It was founded by Dr. Thomas Kelly, a former vascular surgical trainee, with Waleed Mussa and Yu Liu, and is an ambient scribe now growing into an “AI Care Partner.”
What Heidi builds, and why it shapes the loop
Heidi captures audio from a consult, transcribes it, and uses a language model to produce a structured clinical document the clinician reviews and signs. That pipeline is the syllabus: near-real-time audio handling, speech recognition across accents and specialties, LLM summarization that cannot invent a medication or a dose, and a human sign-off step the system is built around. Heidi’s September 2026 messaging adds “supervised agentic workflows”: agents that do more of the work around a visit while a clinician stays in the loop. Every engineering round comes back to keeping a model useful without letting it put a wrong fact in a medical record.
In September 2026 Heidi announced roughly $340 million in new financing in two parts: a $100 million Series C led by Blackbird at about a $900 million valuation, and a separate $240 million growth facility from General Catalyst’s Customer Value Fund, which is go-to-market financing rather than equity (MobiHealthNews, HLTH, DealStreetAsia, September 22 to 23). That roughly doubles the $465 million valuation from its $65 million Series B in October 2025. Heidi says it handles about 2.8 million patient visits a week across around 190 countries (one outlet reported 116), so treat the reach as company-stated. Either way, a fast-scaling startup that rewards people who ship.
Which track, and how junior can you be
Heidi’s open reqs skew mid-to-senior: senior roles ask for five-plus years, and integrations work assumes you have shipped against a real EHR. The clearest junior door is the engineering internship, so early-career candidates and students should target that rather than a senior req. A bootcamp graduate or non-CS career changer should aim at general full-stack product roles, where the screen rewards maintainable code over healthcare experience, then move into integrations later. If the jargon is new: FHIR and HL7 are the standard formats health systems use to exchange data, Epic and Cerner are the dominant EHR vendors, and a BAA (business associate agreement) is the contract that lets a vendor handle patient data under HIPAA.
The loop, inferred from medical-scribe and full-stack norms
Be precise about provenance. Heidi’s job postings establish the skills the roles need, not the interview process; the only posting that speaks to process is the Built In engineering internship, which describes a hiring-manager conversation with a short collaborative design discussion. So the loop below generalizes that intern signal plus medical-scribe and vertical-SaaS norms to senior full-time roles; none of it is confirmed by Heidi. A company this size, hiring through Ashby and the usual boards, almost always runs a compact loop of about five stages over two to four weeks: recruiter screen, technical screen, hiring-manager conversation, a technical round or build, and a final values-and-fit round.
The closest peer is the ambient clinical-AI world Heidi competes in. This site’s Abridge interview guide is the most direct comparison: the same consult-to-note problem, and a loop that leans on ML evaluation, exactly the muscle Heidi’s applied-AI rounds test. Harvey AI is a cross-vertical analogy: AI inside a regulated profession, screening for judgment about where a model may be wrong. For the broader pattern, the AI-native company guides and the AI-startup interview difficulty index map how these loops run, and the AI-era interview guide covers how coding-with-AI shows up in the screen itself.
Heidi Health interview loop at a glance
| Stage (inferred from medical-scribe and full-stack norms and Heidi postings, Oct 2026; no published loop) | What to expect | How to prep |
|---|---|---|
| Recruiter screen | Role fit, motivation, and logistics; often a few quick questions about your most relevant shipped work | State your strongest full-stack or applied-AI project in two sentences, with the impact numbers |
| Technical screen | Coding exercise for product roles, or a systems and past-project discussion for senior and integrations roles | Warm up data structures and one real system you designed; reason about correctness first, speed second |
| Hiring-manager conversation | Domain fit, how you make trade-offs, and a past project the manager pushes on | Bring one decision you made with incomplete information and what you changed when it went wrong |
| Technical round or build | A take-home, a live build, or an architecture discussion grounded in Heidi’s actual problems | Practice designing the audio-to-note pipeline with a human review step; name the failure modes |
| Values and cross-functional fit | Working style, clinician empathy, and how you handle ambiguity and speed | Have a real reason you want to work on clinical documentation specifically |
EHR and EMR integrations: the most concrete track
The most specific engineering title in Heidi’s reqs is EHR and EMR integrations, where a lot of the hard work lives. A scribe is only useful if the note lands back in the system the clinic runs on, so expect FHIR and HL7, reading and writing Epic and Cerner-class systems, each with a slightly different, often legacy, interface. Be ready on idempotent writes, reconciling records when a partner returns partial or malformed data, auth and token refresh across many tenants, and testing an integration you cannot poke freely in production. This is also where SQL and data-modeling questions surface, since clinical data is relational, messy, and audited. Prompts to rehearse:
- A FHIR write returns a partial success: the encounter posted but the note attachment failed. Walk through an idempotent retry that will not double-post the encounter.
- Two tenants on the same Epic version return subtly different field shapes for one resource. How do you detect the drift and keep a single integration codebase?
Applied AI: accuracy is the whole job
For applied-AI roles, the screen is whether you grasp that a confident wrong note is worse than no note. Expect to reason about speech recognition across accents, specialties, and noisy clinic audio, and LLM summarization that must not fabricate a symptom, a medication, or a dose. Give “measure hallucination” teeth: score each note for faithfulness against the source transcript so every claim traces to something said, compare against clinician-written reference notes, and stratify eval sets by clinical category so a gain in GP notes does not hide a regression in cardiology. In supervised agentic workflows, what an agent may do autonomously versus what needs a human gate is itself a design problem, especially for a referral or a billing code. Prompts to rehearse:
- The scribe sometimes adds a medication the clinician never mentioned. Design the evaluation that catches it before shipping.
- Define a faithfulness metric for a note against its transcript, then say what you do about a note that is faithful but omits a key finding.
Full-stack and coding: solid, not competitive
For general product and full-stack roles, plan for a working-level coding round. Refresh the coding patterns catalog and keep your time and space complexity reasoning automatic, then stop; a scaling health-tech startup screens for correct, maintainable code and edge-case reasoning, not contest tricks. The system design guides are worth a pass; one question to rehearse end to end is the audio-to-note pipeline. A workable answer: stream audio in chunks to an ASR service, buffer partial transcripts, hand the finished transcript to an LLM that drafts a structured note, write the draft to a review queue, and require clinician sign-off before anything reaches the record. The parts worth probing: backpressure when ASR lags, retrying note generation without duplicates, and keeping latency low enough for mid-visit use.
Compliance is not a side quiz
Because almost everything touches protected health information, data handling comes up across tracks, well beyond a single security round. Heidi states that it adheres to HIPAA, GDPR, the UK’s NHS requirements, and the Australian Privacy Principles, and that consultation audio is not stored after transcription; these are self-reported, so verify for your market and plan. For prep, know what a BAA is, encryption in transit and at rest, access controls and audit logging, consent for recording, and de-identified versus identifiable data. You need not be a compliance lawyer, but you must show you will not treat PHI like ordinary application data.
Behavioral: clinician empathy and a real why-Heidi
Behavioral rounds here check two things: whether you own hard outcomes, and whether you care about the user, a time-starved clinician. Prepare a project you drove through a real setback; the STAR method keeps it from sprawling, but what lands is the decision you made under uncertainty and what you learned when it broke. The people interviewing you chose healthcare, where a wrong output lands in a patient record, so “I want to work in AI” with no healthcare hook tends to land flat. Prompts worth a rehearsed answer:
- Why clinical documentation specifically, and not any other AI product?
- Tell me about something you shipped that broke in a way that mattered, and what you changed.
Compensation, with the sourcing attached
There is no clean total-comp dataset for Heidi. The one US figure I found is a Senior Integrations Engineer role (New York or remote) showing roughly $160,000 to $220,000 on arc.dev as of October 2026. Treat it as a listed range, not a confirmed employer band: arc mixes employer figures with its own estimates and this one is not clearly marked, though it fits New York pay-transparency norms. A separate listing tagged a San Francisco role in Australian dollars, almost certainly a Melbourne band mislabeled as US; the Dublin and London listings do not disclose salary; and levels.fyi, Glassdoor, and Indeed have no usable Heidi sample.
So treat that range as the one data point, scale by location (Melbourne roles are in AUD, not comparable), and assume equity is part of the package at a company that repriced near a $900 million valuation. Model your own number with the total comp calculator, and negotiate any offer on the equity grant with a real salary negotiation, where a startup at this stage has the most room. Heidi’s problem is sharp: give a clinician their attention back without putting a wrong word in the chart. Candidates who can discuss where the model helps and where it must defer read as people who belong here.
Practice the behavioral round:
