The Startup Teaching AI to Read a Doctor's Bad Handwriting

Roughly 80% of a patient's medical story is unreadable to a computer. It hides in scanned faxes, handwritten notes, discharge summaries and PDFs that pile up faster than anyone can review them. For decades, healthcare's answer was to hire more people to squint at the pile. Keebler Health, a Durham, North Carolina company founded in 2023, has a different answer: teach a large language model to read the whole thing.

The problem the company chose sounds almost aggressively boring - "HCC coding from unstructured clinical data" - and that is exactly the point. Behind that phrase sits one of the largest and least-glamorous markets in medicine: risk adjustment, the machinery that decides how much value-based care organizations get paid to keep patients healthy. Get the diagnoses documented accurately and the payment reflects how sick the population really is. Miss them, and providers eat the cost of care they were never credited for.

01 / THE PROBLEMThe 41% that slips through

Here is the number that keeps risk-bearing providers up at night: across different electronic health record sources, only about 59.4% of chronic conditions get captured consistently. The other 41% is not fraud or laziness. It is a data problem. A patient's diabetes is documented in a cardiologist's scanned note from two years ago; their chronic kidney disease is buried on page seven of a faxed hospital discharge. No human coder has time to read every page of every chart for every patient, so tools sample. Sampling misses things.

That last figure is the one that made investors lean in. In an early pilot on just 44 complex patients, Keebler's platform surfaced 98 previously undocumented conditions. Scale that instinct across a full panel and the math changes: the company says customers routinely identify 15-30% more disease burden than traditional methods, and one provider group had a full 7,000-patient panel risk-adjusted in a fraction of the months it would have taken by hand.

02 / HOW IT WORKSFrom a fax to a flagged diagnosis

Most "AI for healthcare" tools quietly do the easy thing: they touch clean, structured data and leave the messy 80% alone. Keebler was built LLM-native from the ground up to go straight at the mess. It ingests the full chart - structured records and unstructured documents alike - reads them the way a clinician would, and flags suspected conditions before the visit, not months later at billing time.

The last step is the quiet differentiator. It is easy to build AI that guesses a diagnosis; it is hard to build AI that will hand you the page, line and phrase it read to get there. Keebler's outputs are source-linked, which matters enormously when a Risk Adjustment Data Validation (RADV) auditor comes asking why a condition was coded. The feature is not the flashiest thing on a demo. It may be the most important.

03 / WHO BUYS ITThe people paid to keep you healthy

Keebler sells to the organizations that carry risk: value-based care groups, ACOs, health plans, and population health teams. The pitch lands because these buyers are structurally motivated - under value-based contracts, undocumented disease burden is money left on the table and, worse, care that goes unmanaged.

Named customers give the story teeth. Bloom Healthcare reports it doubled its chart-review capacity while staying lean on staffing. Longevity Health Plan says it improved chronic disease identification and documentation as its ACO grew. The company reports processing thousands of patient records weekly, and one customer put its first-year return at 75x - a figure worth treating as a best case rather than a promise, but a striking one all the same.

04 / THE DIFFERENCESoftware eating a services business

Risk adjustment has historically been a services business - rooms full of human coders retrospectively reviewing charts. Keebler's structural bet is that software reads entire populations for a fraction of the cost. The company claims roughly 70% lower cost than services-based competitors, and its LLM-native architecture means it reads unstructured documents directly rather than retrofitting older natural-language-processing tools onto data they were never built for.

There is a reason the fourth name on the founding team is a physician. Isaac Park (CEO, a Duke computer science grad who once co-founded an innovation studio and taught in Duke's engineering school), Andrew Stickney and Kevin Hill - an AI and neuroscience PhD - built the engine. Terrell Bacchus, MD, joined as founding Chief Medical Officer. The pairing shows up in the product philosophy: keep clinicians in the loop, make every output explainable, augment judgment rather than override it.

05 / THE MONEY$16M to read the rest of medicine

In April 2026, Keebler closed a $16M Series A led by Flare Capital Partners and Sands Capital, bringing total funding to around $23M since the company's founding. The round followed an oversubscribed seed - part of $7.8M raised earlier - and is earmarked for commercial growth, team expansion, and a push into adjacent workflows: compliance, population health, and AI-enabled RADV audit readiness.

Yellow chips denote lead investors on the Series A.

The timing is not an accident. CMS has signaled it wants 100% of Medicare beneficiaries in value-based models by 2030 - which means someone, or something, has to read every chart. Keebler pegs its immediate market at $10-15 billion and the broader opportunity, spanning population health, quality measurement and care coordination, at $80-100 billion.

06 / THE TAKEAWAYWhat you can steal from it

The most copyable idea here is not the AI. It is the target. Keebler went after the ugliest, least-defensible data in an entire industry - the faxes and handwriting everyone else routed around - and treated "unreadable" as the moat rather than the obstacle. When the hard input is also the valuable one, being the company willing to read it is the whole business.

The honest caveats are worth stating too. The eye-catching numbers - 75x ROI, 98 findings in 44 patients - come from the company and early pilots, so they read as best cases, not guarantees. LLMs reading clinical documents raise real questions about accuracy and hallucination, which is precisely why the source-linking and clinician-in-the-loop design matter. And this only works where documentation genuinely under-captures disease burden and providers are paid for accuracy - the value-based world. In a fee-for-service clinic with clean data, the pitch loses most of its force. Keebler's bet is that the whole system is moving its way. So far, the checks agree.