Project Lumos
Project Lumos is building a continuous-time, multimodal foundation model of human physiology that learns from raw biometric signals rather than compressed summary metrics. The model uses modality-specific encoders feeding into a shared backbone, enabling new sensors to integrate without retraining. It aims to serve as the missing computational primitive for drug development and clinical monitoring.
- Artificial Intelligence
- Healthcare Technology
- Drug Discovery & Therapeutics
- Digital Health
- Data & Analytics
- Software Only
Funding
Funding not disclosed
Founders
Founder details are not available yet.
Product
Problem
Medicine has never experienced the computational revolution that transformed supply chains and telecommunications. Continuous physiological data—such as a 250-readings-per-second ECG—is reduced to a single threshold-checked number, discarding the vast majority of signal. Drug development has also gotten harder over time, with costs rising from $180M to $2.6B per drug since 1980 while approvals per dollar have declined, and most drugs entering trials fail late.
Solution
Lumos is building a foundation model of human physiology—a continuous-time, multimodal model that learns how the whole body works from raw physiological signals. Instead of training on already-recorded categories, the model trains on full-resolution biometric streams and learns how the body actually changes and responds to disease, drugs, and intervention. Each sensing modality gets its own encoder, and a shared backbone maps them into one continuous, evolving representation of a patient's state. This architecture allows new sensors to plug in without retraining while the model keeps learning. The goal is to make physiology computable, serving as the missing primitive beneath modern medicine and drug development.
Target Audience
Lumos targets pharmaceutical companies facing rising R&D costs and clinical failure rates, as well as healthcare providers and researchers who need continuous, quantitative understanding of disease progression and drug response from physiological data.
Features
- Continuous-time, multimodal foundation model trained on raw physiological signals rather than compressed summary metrics
- Modality-specific encoders feeding into a shared backbone that produces an evolving representation of patient state
- Ability to integrate new sensing modalities without model retraining
- Learns from full-resolution streams including ECG (250 readings per second), voice biomarkers, movement, and sleep data
- Designed to capture subtle physiological changes—such as micro-tremors in voice that can flag Parkinson's progression and depression