Replica Health

Replica Health is an American artificial-intelligence research lab and software company headquartered in New York. It describes itself as building and shipping software for chronic conditions that the technology industry has passed over, deliberately targeting patient populations that are too small for big tech to serve. Its stated mission is to bring state-of-the-art AI to those patients as quickly as possible. The company operates through a repeatable playbook. First, it identifies a gap: a condition where patients face daily, data-rich problems but where the affected population is small enough that mainstream digital health never shows up. Second, it builds a reusable engine of standardized datasets, high-speed simulators and rigorous validation pipelines. Third, it ships fast, using AI-accelerated R&D and on-device machine learning to move from idea to working product in months. Replica Health's flagship program is in metabolic health, specifically Type 1 diabetes. It has shipped an iOS app, developed patent-pending carbohydrate-absorption modeling, assembled what it calls the largest standardized T1D dataset, and built an algorithm pipeline aimed at the next generation of automated insulin delivery. A newer program, Voice DeepLarynx, provides real-time voice reconstruction for people who speak with an electrolarynx after laryngectomy, producing natural speech in the user's own voice and running entirely on the device. Beyond its own programs, Replica Health positions its AI/ML stack as available to partners. The stack includes MetaboNet, a published, standardized Type 1 diabetes dataset described as the "ImageNet for diabetes," with 2,030 patient-years across 3,135 subjects and 21 source datasets and tiered public and DUA-governed access; MetaboNet-Bench, a multi-modal glucose-forecasting benchmark developed with Stanford; CarbCast, a patent-pending on-device model that predicts the full carbohydrate-absorption curve from a meal description and context; EventSense, an on-device model that recognizes meals and activities from CGM trend, time of day, location and activity; RapidUVAPadova, a re-engineered, massively parallel UVA/Padova simulator; MetaboNet-Replay, a sim-to-real validation and counterfactual engine; and Replica-Capture, an iOS app and data backend for real-world data collection in clinical trials. The company emphasizes open research, publishing work intended to move the whole field forward, and a preference for models that run on the user's hardware rather than in the cloud. It is a private company; no parent, owner or listing information is publicly indicated in the available evidence.

Replica Health is a New York-based company that presents itself as an AI research lab rather than a conventional digital-health vendor. Its founding premise, as stated on its own site, is that the tech industry has passed over many chronic conditions, and that patient populations can be too small for big tech to serve. The company frames its work as finding those gaps and bringing state-of-the-art AI to them as quickly as possible. Rather than pursuing one product, Replica Health developed a repeatable playbook. It looks for conditions where patients face daily, data-rich problems but where the population is small enough that mainstream digital health never shows up. It then builds a reusable engine — standardized datasets, high-speed simulators and rigorous validation pipelines — and ships fast using AI-accelerated R&D and on-device machine learning, aiming to go from idea to working product in months. Its flagship program is in metabolic health, specifically Type 1 diabetes. The company reports having shipped an iOS app, developed patent-pending carbohydrate-absorption modeling, assembled what it describes as the largest standardized T1D dataset, and built an algorithm pipeline for the next generation of automated insulin delivery. A newer program, Voice DeepLarynx, addresses a very different population: people who speak with an electrolarynx after laryngectomy. It provides real-time voice reconstruction that produces natural speech in the user's own voice, running entirely on the device. Alongside its own programs, Replica Health has built an AI/ML stack that it makes available to partners. The published MetaboNet dataset — described as the "ImageNet for diabetes" — consolidates 2,030 patient-years across 3,135 subjects and 21 source datasets, with tiered public and DUA-governed access at metabo-net.org. MetaboNet-Bench, a multi-modal glucose-forecasting benchmark, was developed in collaboration with Stanford. CarbCast is a patent-pending on-device model that predicts the full carbohydrate-absorption curve from a meal description and context, in the format insulin-delivery systems accept; it was evaluated head-to-head against human carb estimates in the Tidepool × Stanford study presented at ADA 2026. EventSense recognizes meals and activities from on-device context such as CGM trend, time of day, location and activity, enabling one-tap logging with over-the-air model updates. RapidUVAPadova is a re-engineered, massively parallel UVA/Padova simulator reported to be 14× faster for in-silico testing across virtual patients, with a Python interface for new control policies including multi-GPU workloads. MetaboNet-Replay reconstructs real patient trajectories from MetaboNet and replays them under new insulin-delivery policies, acting as a counterfactual engine to optimize and de-risk AID algorithms per patient. Replica-Capture is an iOS app plus data backend that can power real-world data collection for clinical trials, capturing diabetes, activity and location data. The company emphasizes open research, stating that it publishes work intended to move the whole field forward, and a preference for models that run on the user's hardware rather than in the cloud. It also signals that more initiatives are to come, evaluating new places where continuous sensing and on-device AI can improve people's lives. The available evidence does not disclose founding year, founders, funding, ownership or listing status; it is a private company headquartered in New York.

• MetaboNet published as a standardized Type 1 diabetes dataset with tiered public and DUA-governed access at metabo-net.org.

• MetaboNet-Bench multi-modal glucose-forecasting benchmark developed in collaboration with Stanford.

• CarbCast evaluated head-to-head against human carb estimates in the Tidepool × Stanford study presented at ADA 2026.

• Voice DeepLarynx introduced as a new program for real-time voice reconstruction after laryngectomy.

An AI research lab bringing frontier AI to overlooked patient populations, with a reusable engine of datasets, simulators and on-device models available to partners.

• Type 1 Diabetes program

  Flagship metabolic health program: a shipped iOS app, patent-pending carb-absorption modeling, the largest standardized T1D dataset, and an algorithm pipeline for the next generation of automated insulin delivery.

• Voice DeepLarynx

  Real-time voice reconstruction for people who speak with an electrolarynx after laryngectomy — natural speech, in the user's own voice, running entirely on the device.

• MetaboNet

  The "ImageNet for diabetes" — the largest standardized, consolidated Type 1 diabetes dataset: 2,030 patient-years across 3,135 subjects and 21 source datasets, with tiered public and DUA-governed access at metabo-net.org.

• MetaboNet-Bench

  A multi-modal glucose-forecasting benchmark developed in collaboration with Stanford.

• CarbCast

  Patent-pending on-device model that predicts the full carbohydrate-absorption curve from a meal description and context, in the format insulin-delivery systems accept.

• EventSense

  On-device model that recognizes meals and activities from on-device context — CGM trend, time of day, location, and activity — enabling one-tap logging, with over-the-air model updates.

• RapidUVAPadova

  A re-engineered, massively parallel UVA/Padova simulator — 14× faster in-silico testing across virtual patients, with a Python interface for new control policies, including multi-GPU workloads.

• MetaboNet-Replay

  Reconstructs real patient trajectories from MetaboNet, then replays them under new insulin-delivery policies — a counterfactual engine that can optimize and de-risk AID algorithms on a per-patient basis.

• Replica-Capture

  An iOS app plus data backend that can power real-world data collection for clinical trials — capturing diabetes, activity, and location data.

1. 1Replica Health — Frontier AI for overlooked patientsreplica.health
2. 2Platform & Services | Replica Healthreplica.health
3. 3Replica Health (United States) — Wikidatawikidata.org