HelixCare Labs

HelixCare Labs occupies one floor of a Kendall Square office building wedged between two larger biotechs, a Korean barbecue restaurant, and a pharmacy. Eleven people, a managing director from the hospital systems world, and a lead engineer who spent eight years at a major imaging startup before deciding she preferred companies whose names she could fit on a business card.

Their product line is diagnostic-support models — radiology second-opinion, dermatology screening, the kind of clinical AI that needs to fail in clinically defensible ways. Models ship for low to mid six-figure Cu, depending on clinical scope; the firm invoices through hospital procurement systems that occasionally take longer to process the payment than the firm took to deliver the model. They take CE marking seriously. They take FDA submissions seriously. They take audit trails so seriously that their analysts have a running internal joke about being able to reconstruct any past model’s training history from memory, on a whiteboard, drunk.

They have never tested this claim and decline to.

HelixCare’s archival policy mandates retention of every model version, every training corpus, every weight checkpoint, indefinitely. The managing director, when asked why, gives a long answer about regulatory exposure. The actual reason — known only inside the firm — is that one of their early models flagged a pattern in an old radiology dataset that, when investigated, turned out to correspond to nothing the firm has been willing to commit to writing.