Healthcare IT · Hospital quality and patient safety
Pharos
An AI system that reads hospital charts and produces the quality measures nurses now extract by hand. Seed funded in October 2024, tested against human abstractors on about 2,000 pathology reports, and still waiting on a published paper.
Eight hours for one patient
A nurse sits down with one hospital admission. Not to treat the patient, who went home weeks ago, but to read the chart: triage note, operative report, pathology, discharge summary. Buried in those pages are the facts a clinical registry wants, in the exact form it defines them.
“A single case can take up to eight hours” to report, Ryan Isono of Felicis told TechCrunch in 2024. That is what a complex case costs.
Multiply it. The American College of Surgeons tells hospitals one full-time surgical clinical reviewer covers about 1,680 cases a year. Pharos says a single hospital can spend $5 million a year on this, and the data still lands weeks after discharge, on a sample rather than the whole population. That is the seam it is trying to open: not a new clinical model, but the machinery that decides whether a hospital ever finds out what went wrong.
Why the measurement is the problem
The scale of the harm is a matter of public record. In 2022 the HHS Office of Inspector General reviewed the charts of 770 Medicare patients discharged in October 2018 and reported that “Twenty-five percent of Medicare patients experienced patient harm during their hospital stays in October 2018.” Twelve percent had adverse events, meaning longer stays, permanent harm, life-saving intervention or death. Physician reviewers judged 43 percent of all harm events preventable.
The OIG did that work the way hospitals do: nurses screened, then physicians read. Accurate, and it does not scale. That is the argument for automating abstraction, and also why the accuracy bar is unforgiving. A wrong measure is worse than none, because it points a quality improvement project at the wrong target.
The pressure keeps rising. CMS moved the SEP-1 sepsis bundle into Hospital Value-Based Purchasing, so sepsis process adherence now touches payment. Surgical, stroke, pathology and trauma registries each want their own variables, defined their own way. Every new registry is another full-time reviewer, or another gap.
Two engineers, one doctor, and a sepsis model
Felix Brann and Matthew Jones met at Vital, which puts software into emergency departments. Over nearly five years they deployed predictive AI into more than 70 hospitals, integrating with every major electronic health record vendor. Brann was VP of Data Science there, and before that VP of Quantitative Research at J.P. Morgan.
His research record is checkable. In January 2024 he was first author on a JMIR AI paper predicting sepsis at emergency department triage from nursing notes, across a retrospective cohort of 1,059,386 adult encounters at four emergency departments, reporting an area under the curve of 0.94 at triage. A 2020 JACEP Open paper applied similar methods to resource prediction. Those two are the only papers indexed under his name in PubMed.
The sepsis model is where the company started. Piloting it, the two concluded that the people who most needed AI were not the bedside clinicians but the quality staff behind them. They entered Y Combinator in Summer 2024 and added a third co-founder mid-batch: Alex Clarke, a doctor who trained in the UK, took a PhD in artificial intelligence at Imperial College London, and researched medical AI at Meta. He had spent his residency abstracting by hand.
“We believe in a future where AI catches medical mistakes everywhere in the hospital, before they become serious. We want to be the lighthouse for our hospitals, supporting clinicians and reducing patient harm.” That is the founders on their own company page, and it explains the name.
A conference stage, and a paper that has not arrived
On October 25, 2024 Pharos announced an oversubscribed $5 million seed led by Felicis, with General Catalyst, Moxxie and Y Combinator. Moxxie had led the pre-seed. The company was then three people.
Thirteen months later came the result that matters. In November 2025 Pharos presented at the Council of Medical Specialty Societies annual meeting, a room run by Helen Burstin, formerly chief scientific officer of the National Quality Forum. The work used College of American Pathologists measures, with gold labels from abstractors at Kiser Healthcare Solutions. Brann posted the headline himself: the AI matched two skilled human abstractors on accuracy and inter-rater reliability across roughly 2,000 double-abstracted records.
He also wrote five words that are still true ten months on: “Stay tuned for the paper”.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Abstraction accuracy | Presented at the CMSS annual meeting, November 2025. Four College of American Pathologists measures, about 2,000 pathology reports, gold labels from abstractors at Kiser Healthcare Solutions. Company account: matched human accuracy and inter-rater reliability. Not peer reviewed, not replicated, and limited to pathology reports rather than whole patient journeys. | Company-stated |
| Peer reviewed paper on the product | None. A PubMed author search on September 22, 2026 returns two papers for Felix Brann, both from his Vital years, neither on the Pharos product. | Not found |
| Founder research record | JMIR AI, Jan 2024. Retrospective cohort of 1,059,386 adult encounters, four emergency departments. Area under the curve 0.94 at triage, sensitivity 0.87, specificity 0.85, positive predictive value 0.18. | Public record |
| FDA record | No submission by this company. The five 510(k) clearances under a Pharos name, 2007 to 2011, belong to Pharos Life Corporation, for Tanda skincare devices. | Public record |
| Trials and federal grants | None for this company. ClinicalTrials.gov records naming Pharos are sponsored by Pharos University in Alexandria, Egypt. A 2004 NIH award to Pharos, LLC funded an artificial lung. | Public record |
| Named hospital customers | None found in any public source as of September 22, 2026. Job listings cite hospital partners without naming one. | Not found |
| Security attestation | The website shows a SOC2 Type 2 badge and says “Our HIPAA and Soc 2 Type II compliant platform runs in the cloud using modern security practices.” Its security page, updated September 2026, says “Pharos is fully HIPAA compliant and currently undergoing SOC2 Type 2 audit certification”. Those do not agree. | Claims disagree |
| Funding | $5,000,000 seed announced October 25, 2024. No Form D for this company appears in SEC EDGAR as of September 22, 2026. | Public record |
Read plainly: the strongest thing Pharos has shown is a head-to-head accuracy result on a narrow, well-chosen slice of the problem, presented to the right audience with the right collaborators, and not yet written up. A real result, and a thin one. Pathology reports are messy narrative text, which makes the test hard, but they are short and self-contained next to a full inpatient stay. Nothing public shows the system working on a surgical registry, on a whole admission, or inside a named hospital. The comparison on rigour is Bayesian Health, which took the slower path: a prospective study in Nature in 2022, then FDA clearance in May 2026. Pharos sells administrative software, not a diagnostic, so it needs no clearance. It does need the paper.
What to watch
- The CMSS paper. Promised November 2025, unpublished as of September 22, 2026.
- A named hospital customer, or a registry partner beyond the CAP work.
- Results on a full inpatient record, not a pathology report, which is the broader marketing claim.
- A completed SOC 2 Type 2 report, settling the contradiction on the company website.
- A Series A, or a first Form D. Neither exists 22 months after the seed.
In their words
“A single case can take up to eight hours”
Ryan Isono, Felicis, TechCrunch, 2024 · Independent
“This manual work is expensive, pulls clinical staff away from the front-line, and means that hospital teams get data long after patients are discharged.”
Felix Brann, CEO, seed announcement, 2024 · Company release
“We give teams data in real-time, letting them take action on process gaps before they turn into patient harm.”
Felix Brann, seed announcement, 2024 · Company release
“The mundane areas of healthcare administration make up nearly $1 trillion in spend annually.”
Ryan Isono, Felicis, seed announcement, 2024 · Company release
“That Venn diagram doesn’t normally overlap. That’s why we think we’re going to win.”
Felix Brann, TechCrunch, 2024 · Independent
“Pharos (YC S24) AI had the same accuracy and IRR as two skilled human abstractors, across 2000 independently double-abstracted records.”
Felix Brann, LinkedIn, November 2025 · Company-stated
“Across four CAP measures (about 2,000 reports), the system matched human performance, with low-variance outputs and fully reviewable reasoning for each decision”
Pharos, LinkedIn, 2025 · Company-stated
“Pharos will need to prove its accuracy matches or exceeds human abstractors to win trust from quality teams.”
Dealroom, 2024 · Independent
Related companies
Sources
- Public recordAdverse Events in Hospitals: A Quarter of Medicare Patients Experienced Harm
- Public recordSepsis Prediction at Emergency Department Triage Using NLP
- Public recordPrediction of ED resource requirements during triage
- Public recordPubMed author search, Brann F
- Public recordEDGAR full text and company search, form D
- Public record510(k) clearances, Pharos Life Corporation
- Public recordACS NSQIP FAQ, reviewer caseload
- Public record2025 Pathology Quality Measures
- IndependentYC startup Pharos lands a $5M seed led by Felicis
- IndependentPharos Raises $5M to Automate Hospital Quality Reporting
- IndependentPharos raises $5M Seed, deal analysis
- IndependentBayesian Health gets FDA nod for AI sepsis detection tool
- IndependentSEP-1 and Hospital Value-Based Purchasing
- IndependentHelen Burstin, chief executive officer
- CompanyPharos Announces $5M Seed Round
- CompanyLaunch YC: Pharos, and company page
- CompanyWebsite copy, security and privacy pages
- CompanyBrann and Pharos posts on CMSS 2025 results
- CompanyPharos open roles, three listings
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
