For a teenager being treated for anorexia or bulimia, the riskiest hours are rarely the ones spent inside a clinician's office. They are the evenings at the dinner table, the late-night scroll, the weekend when a parent is trying to remember what the family-based therapist said to do if a meal gets refused. Kuddo, an early-stage San Francisco healthtech company, is building software for exactly that gap: an AI companion meant to extend eating disorder care beyond the 50-minute session and give clinicians a way to measure whether the therapy actually being delivered matches the protocol on paper.
The company describes itself as an "AI-driven platform for patient insights in therapy settings" [websets.exa.ai], with an explicit focus on eating disorders [LinkedIn]. On its site, Kuddo frames the product as an "AI Companion for Eating Disorder Treatment Fidelity" aimed at "Scaling Quality, Not Just Access" [LinkedIn, 2026]. That framing matters. Most digital mental health entrants over the past decade have sold access: more therapists, faster matching, cheaper sessions. Kuddo is making a narrower bet, that the bottleneck in eating disorder care is not only supply of clinicians but the consistency and adherence of the care they deliver.
The bet
Kuddo's technical wedge, as the company describes it, is a set of fine-tuned large language models that "identify therapist behaviors, treatment phases, and adherence" [kuddo.club]. In practice, that points toward session analysis: ingesting therapy interactions and scoring them against an evidence-based protocol such as Family-Based Treatment or Enhanced Cognitive Behavioral Therapy, the two manualized approaches that dominate the field. The company also says its system is designed to "detect high-risk behaviors early and intervene promptly to ensure patient safety and well-being" [kuddo.club] and to provide "continuous support for families dealing with eating disorders, ensuring proper guidance even when the therapist isn't" [kuddo.club].
The buyer, given the B2B model, is most plausibly a treatment program, an academic clinic, or a payer-aligned provider group rather than the patient directly. Kuddo argues a compounding advantage on that side: "Benchmarking improves with each partner site, quality moat grows over time" [kuddo.club]. If that holds, every additional clinic that contributes session data sharpens the fidelity scoring for the next one.
Why the patient population matters
Eating disorders carry one of the highest mortality rates in psychiatry, and the standard of care is unusually well defined but unevenly delivered. For adolescents with anorexia nervosa, Family-Based Treatment is the first-line recommendation across most clinical guidelines, typically delivered in roughly 20 sessions over a year, with weight restoration handled at home under parental supervision and medical monitoring by a pediatrician or adolescent medicine specialist. For adults, Enhanced Cognitive Behavioral Therapy and, in more acute cases, residential or partial hospitalization programs form the backbone. The persistent problem is fidelity drift: studies have long shown that clinicians trained in manualized protocols gradually deviate from them, and that families, once discharged from intensive programs, often lose the scaffolding that made treatment work. There is no FDA-cleared digital therapeutic specifically indicated for eating disorders at the level of a prescription product, though the broader category of prescription digital therapeutics in mental health has seen FDA authorizations in adjacent conditions.
That is the seam Kuddo is trying to sit in. Not replacing the therapist, not promising a diagnostic, but instrumenting what already happens and supporting the family in between.
The team
Kuddo was founded by Wenyi Zhu, who serves as Co-Founder and CEO [websets.exa.ai] and was educated at Carnegie Mellon University [LinkedIn], including the Tepper School of Business as a Tepper ACS Scholar [LinkedIn]. Zhu's prior roles span CyberConnect, Protocol Labs, Citi and CICC [rocketreach.co], a background weighted toward fintech and crypto infrastructure rather than clinical care. The clinical credibility on the cap table comes from Anne Claire Grammer, PhD, who is listed as Chief Clinical Officer [LinkedIn, 2026] and is described as a licensed clinical psychologist focused on AI and youth mental health, with "expertise in training methodologies and clinical experience across academic and industry sectors" [zoominfo.com]. The pairing of an operator-founder with a clinical psychologist who specializes in training is consistent with the fidelity-scoring thesis: the product is, in effect, an automation of what a clinical supervisor would otherwise do by hand.
The honest counterfactual
The bear case is real and worth naming. Eating disorder care is one of the most clinically and ethically sensitive corners of behavioral health. An AI companion that misreads a high-risk behavior, or that a family over-relies on in place of escalation to a clinician, is a patient safety problem, not a product bug. Regulators have been increasingly explicit that software making clinical claims, including risk detection, can fall under FDA oversight as Software as a Medical Device, and the agency has flagged generative AI in mental health as an area of active scrutiny. What bulls would answer is that Kuddo's stated framing is supportive rather than diagnostic, with the therapist and family kept in the loop, and that fidelity measurement, the company's other core use case, is closer to a clinical operations tool than a regulated device. The honest read is that the regulatory posture will depend heavily on how the product is marketed and what claims accompany the high-risk behavior detection feature [kuddo.club]. Peer-reviewed evidence on the fidelity-scoring models has not yet been published in the sources reviewed, and that, more than anything, is what would move this from interesting to validated.
What to watch
The next twelve months should clarify three things. First, the identity of Kuddo's first named partner sites: an academic eating disorder program or an established residential provider would be a strong signal, both for distribution and for the data flywheel the company is describing. Second, any peer-reviewed or conference work from Grammer and the team validating that the fine-tuned models actually agree with human supervisors on fidelity coding, which is the standard test in the implementation science literature. Third, the regulatory framing: whether Kuddo positions the high-risk behavior feature as decision support inside a clinician's workflow or as something families interact with directly, because those are very different conversations with the FDA.
Disease state: eating disorders, principally anorexia nervosa and bulimia nervosa, with an apparent emphasis on adolescent and family-based care. Patient population: patients in active outpatient or step-down treatment and the families supporting them. That is a small, high-acuity, under-served group, and a thoughtful tool built specifically for them is, on its face, a worthwhile thing to try.
Pulse Raman, Health and Bio Correspondent.