Daniel Nadler, who founded Kensho and sold it to S&P Global in 2018, started OpenEvidence with Harvard machine learning PhD Zachary Ziegler, now CTO. Most sources put the founding at 2022 (Sacra and some job boards say 2021). The product is a search-and-answer engine over roughly 35 million peer-reviewed publications plus licensed content from the New England Journal of Medicine and JAMA Network, and it answers only when it can cite evidence. It began in Cambridge, Massachusetts, joined Mayo Clinic Platform Accelerate in 2023, and moved its headquarters to Miami in summer 2025. It raised a $75 million Series A at $1 billion from Sequoia in February 2025, $210 million at $3.5 billion (GV, Kleiner Perkins) that July, $200 million at about $6 billion in October, and a $250 million Series D at $12 billion (Thrive Capital, DST Global) in January 2026, roughly $700 million in a year. Sacra estimated annualized revenue near $300 million by July 2026. Micah Smith is VP of Engineering, and the 2026 postings describe an engineering team of about 30.
Process
OpenEvidence does not publish its interview stages, and the public record is thin: three Glassdoor interview reviews and one Blind thread. The company’s Ashby board says more. September 2026 openings include Member of Technical Staff, Software Engineer (Data Platform, Backend and Infrastructure, Applied AI, Site Reliability), Security Engineer, and Research Scientist. Every engineering role except SRE and security is five days a week on-site in San Francisco or Miami; those two also list Remote. The Member of Technical Staff posting says there are no product managers, engineers own projects end to end, and it wants people who are “smart, ambitious, scrappy, autonomous, precise, motivated, hard-working, and low-ego.” The CTO told GV that the hiring question is whether the person is one of the smartest people you have met and whether you would trust them with autonomy. Read that as a loop that weights judgment and ownership over textbook coding drills.
The one stage every candidate report agrees on is a take-home project, and it is heavy. A Glassdoor reviewer for a San Francisco engineering role in August 2023 described a two-week project: pull data from Google Drive and several medical databases, then serve queries over research papers in Python in under eight seconds. Another Glassdoor report describes an assignment wanting two CI/CD pipelines, infrastructure as code, and a real-time load balancing algorithm, and suspects the work was something the company needed built. A Blind poster in November 2025 spent 16 hours on it, emailed the recruiter and tech lead several times, never heard back, then was contacted by ten different recruiters for the same role. Glassdoor’s average process length is 28 days. Cap your take-home at a fixed number of hours, say so in the submission, and get the review timeline in writing first. The system design and behavioral parts of the loop are not documented anywhere public; the shape below is an inference from the postings and from peer loops such as the one in our Abridge interview guide.
| Stage | Who | What it tests |
|---|---|---|
| Recruiter screen (reported) | Recruiter | Background, willingness to work on-site in San Francisco or Miami, interest in clinical AI |
| Take-home project (reported, one to two weeks) | Tech lead, reviewed asynchronously | Building a working retrieval or infrastructure system end to end: ingestion, query latency, deployment, tests |
| Technical review of the take-home (inferred) | Tech lead or CTO | Defending design choices, extending the project live, tradeoffs in latency, correctness, and cost |
| Founder or leadership conversation (inferred) | CTO, VP Engineering, or CEO | Ownership without product managers, judgment under ambiguity, fit with a small on-site team |
What they actually ask
- Design the retrieval layer for a system that answers a clinician’s free-text question from 35 million papers and returns citations. Walk through indexing, ranking, and what you do when the top results disagree with each other.
- The product policy is that the model may not answer without evidence. How would you detect and block an answer whose citations do not support the claim, and how would you measure your miss rate on real physician queries?
- Traffic hit a million consultations in a day in March 2026 on a Python backend on Google Cloud with a Next.js front end on Vercel. Where does that stack break next, and what do you change first?
- The Visits feature transcribes a patient encounter and enriches the note with guidelines. Which parts of that flow touch protected health information, where is it stored, and how do you keep transcripts out of model training and logs?
- Cedars-Sinai, Mount Sinai, and Sutter Health run OpenEvidence inside Epic with patient context from the chart. Design that integration so a query can use the patient’s medications and allergies without the vendor holding the record longer than it must.
- Revenue comes from pharma ads shown next to answers. How do you serve an ad relevant to the clinical question while guaranteeing it has no influence on the ranked evidence, and how would you prove that to a skeptical physician?
Levels and comp (2026)
Titles are flat: Member of Technical Staff or Software Engineer with a specialty suffix, with no senior or staff tiers visible on the careers board. None of the syndicated copies of the postings I read carried a salary band, even for the San Francisco roles, so ask for the range in the first call and check the Ashby listing itself for a California-required band. Public data is scarce: levels.fyi’s OpenEvidence page still showed its submission threshold unmet in September 2026, and Glassdoor has two employee reviews. The best public signal is H-1B disclosure data, which h1bgrader summarizes as a median LCA wage in the low-to-mid $200,000s for fiscal 2026; treat that as a rough floor for base at experienced levels, not a quote. Expect base plus stock options on a four-year vest and possibly a signing bonus.
The equity is the interesting part and the risky part. The last priced round was $12 billion in January 2026; Becker’s reported in July 2026 that the company weighed a $200 million raise at $20 billion and was unlikely to proceed. Roughly $700 million of preferred stock sits ahead of common in a downside exit. Sacra estimates gross margins near 90 percent, unusually good for health AI, but revenue is concentrated in pharma advertising, one buyer category sensitive to regulation. Ask for the current 409A price, your strike, whether early exercise is allowed, whether there has been a tender offer, and how refresh grants work at a company that doubled its valuation twice in a year.
Prep priorities
Learn how physicians use evidence before you touch code. Know the difference between a randomized trial, a meta-analysis, and a specialty society guideline, why a 2019 guideline can outrank a 2024 single-center study, and what GRADE certainty ratings mean, because every ranking question here is really a question about which evidence deserves a citation. Read the DeepConsult, Visits, and Cedars-Sinai announcements and form an opinion on what each needs from the backend. Then sharpen retrieval-augmented generation evaluation (citation recall, faithfulness scoring, physician-rated correctness), Python services at high request volume on GCP, and reliability for a product hospitals depend on during clinical hours. For applied AI and research roles, the AI and ML interview question archive covers evaluation and fine-tuning topics that map to the company’s stated preference for small in-domain models over general-purpose ones.
Before accepting the take-home, ask the recruiter for the expected hour budget, whether a scoped-down submission is acceptable, who reviews it, when you will hear back, and what stages remain. Ask whether on-site is negotiable and get the level and range in the same call. The company’s hiring language, no product managers and full ownership from conception to launch, is the prompt for your behavioral stories: bring two examples where you scoped, built, shipped, and measured something with nobody directing you, and one where you killed your own project because the approach was not working.
Frequently Asked Questions
Does OpenEvidence require a take-home project?
Every public candidate report mentions one, and they are large: a two-week build in the 2023 Glassdoor report, 16 hours in the November 2025 Blind post, and an infrastructure assignment with CI/CD pipelines in another review. Fix your own time budget and confirm the review timeline first; the consistent complaint is slow or absent follow-up.
Is OpenEvidence remote?
Mostly no. The Member of Technical Staff, Data Platform, Backend and Infrastructure, and Applied AI postings require five days a week on-site in San Francisco or Miami. Only the Site Reliability and Platform Security roles list Remote as of September 2026. Headquarters moved from Cambridge to Miami in 2025.
How hard is the OpenEvidence interview compared with other health AI companies?
The stated bar is high: the CTO’s test is whether you are one of the smartest people the interviewer has met and can be trusted with autonomy, on a team of 25 to 30 engineers. Against the loops in our AI startup interview difficulty index, the distinguishing feature is the size of the take-home rather than the number of live rounds.
Is OpenEvidence a scribe company like Abridge?
No. The main product is a search-and-answer engine over the medical literature, free to verified clinicians and funded by pharma advertising. It added an ambient note tool called Visits in August 2025 and a research agent called DeepConsult, and it runs inside Epic at Sutter, Mount Sinai, and Cedars-Sinai, but the retrieval and citation engine is the center of the company and of its interview questions.
Similar company guides
Prepping for OpenEvidence? Put it to work:
