Doses AI develops highly compressed frontier‑scale language and vision models that run entirely on consumer devices, enabling offline inference with no data leaving the device. Their on‑device engine powers privacy‑first health applications such as menstrual cycle prediction, baby‑sleep forecasting, chronic‑condition flare alerts, and medication tracking, as well as nutrition estimation tools. The technology provides real‑time, secure insights for users and developers who require strict data‑sovereignty on smartphones and laptops.
Funding
Funding not disclosed
Founders
Product
Problem
Many health and personal‑data applications rely on cloud‑based AI, requiring users to upload sensitive information to remote servers. This creates privacy risks, adds latency, and limits use in environments with limited connectivity or strict data‑sovereignty requirements.
Solution
Doses AI builds frontier‑scale language and vision models that are heavily compressed (1.58‑bit ternary weights) and can run entirely on the user’s own device, such as a smartphone or consumer laptop. Their custom on‑device inference engine (trellis.cpp) executes mixture‑of‑experts models with a 24 GB on‑disk footprint, delivering offline inference without any data leaving the device. Leveraging this technology, Doses AI offers a suite of health‑focused applications—Ovara for menstrual cycle prediction, Hush for baby‑sleep forecasting, Sana for chronic‑condition flare alerts, and Dosewolf for medication tracking—each powered by on‑device AI that keeps personal health data private while providing real‑time insights. The platform also includes on‑device nutrition estimation and food‑label mapping tools that use agent‑swarm data refinement to improve accuracy without cloud dependence.
Target Audience
Primary users are privacy‑conscious consumers seeking personal health tracking (e.g., menstrual health, infant sleep, chronic condition monitoring) and developers or enterprises building health applications that must keep data on‑device, such as hospitals, banks, or government agencies with strict data‑sovereignty policies.
Features
- 1.58‑bit ternary weight compression enabling 100 B‑parameter mixture‑of‑experts models to fit on consumer hardware (e.g., 24 GB on a MacBook M3)
- Custom inference stack (trellis.cpp) that operates directly on ternary weights for CPU, GPU, and Apple Silicon back‑ends, eliminating floating‑point conversion
- Fully offline operation: no network connection required for inference, ensuring health data never leaves the device
- Suite of privacy‑first health apps (cycle, baby sleep, chronic‑condition flare prediction, medication tracking) built on the same on‑device model family
- On‑device nutrition estimation with a 10‑agent swarm that refines food‑to‑database mappings, reducing calorie‑estimation error by over 60 %
- Compatibility with popular consumer devices (smartphones, laptops) and integration via lightweight SDKs for developers
- Open‑source model components and research publications demonstrating stable ternary mixture‑of‑experts recovery at scale