Komachi AI recommends an alcohol action - abstain, reduce, or maintain, drawn from an optimal treatment policy learned on longitudinal survival data, grounded in cardiovascular-prevention evidence, and checked by a clinical judge before it is shown.
Illustrative sample output, no patient data.
A survival-derived policy proposes the action; an evidence layer grounds it; a clinical judge checks it before anything is shown.
A KL-regularized SAC+BC policy learns the alcohol action that maximizes a survival-derived reward log(1 − λ), anchored to the behavior policy for offline credibility.
See the cohortAn open-weights model, RL-posttrained against the Stage-1 survival reward, reads the individual trajectory and proposes the action in plain language.
Open chatRetrieval over ACC/AHA, ESC, USPSTF, WHO and alcohol-specific guidance surfaces the passages behind each recommendation, with citations.
Open dashboardAn LLM-as-judge scores every recommendation for clinical validity, evidence support, safety, and consistency with the policy before it is displayed.
Open dashboardThe policy is fit on the NHANES I Epidemiologic Follow-up Study - alcohol exposure and mortality tracked across four waves over two decades.
A small group working at the intersection of causal inference, reinforcement learning, and clinical machine learning.
Toru Shirakawa is an MD and PhD student at UC Berkeley and UCSF, working on causal AI for long-term chronic disease management. He develops methods that turn longitudinal health trajectories into reliable, survival-informed treatment policies, bridging clinical medicine and machine learning, and is the lead author of the survival policy-learning framework that Komachi AI is built on.
Sylvia Cheng is a PhD student at UC Berkeley who likes developing ML models and causal methods to tackle real-world problems. Her research is driven by rich interests in applicable mathematical theory, such as dynamic treatment regimes via Targeted Maximum Likelihood Estimation, robust reinforcement-learning policy learning, and evolutionary game theory, which she brings together to make causal estimates trustworthy in practice.
Abhijith Varma Mudunuri is a Senior at UC Berkeley majoring in Data Science and Business. He works in the UC Berkeley / UCSF Department of Computational Precision Health under Prof. Ahmed Alaa on clinical machine learning, computer vision, and reinforcement learning, and is also a researcher with the Berkeley AI Research (BAIR) NLP Group under Prof. Alane Suhr, the Center for Targeted Machine Learning, and the Center for Edge Medicine.