Connectivity resilience
Use a compressed language model on an Android phone to perform inference without an internet connection. The target hardware is a device with at least 6 GB of RAM.
The PocketSage product concept
Useful guidance shouldn't require a perfect connection.
PocketSage has a working Android prototype of an offline-first AI triage tool designed initially for community health workers in low-connectivity settings.
See PocketSage in action
Watch the prototype demo.
A short demonstration of the PocketSage experience. The project is at the prototype stage; clinical validation and field pilots remain ahead.
Watch on YouTube ↗✳ On-device AI · Prototype
The product vision
Designed around the realities of care.
Digital healthcare experiences often assume that a stable connection is always available. PocketSage starts from a different question: how might guidance remain accessible when connectivity is limited?
The prototype uses on-device language-model inference. Planned development includes region-specific clinical content, multilingual and low-literacy workflows, and opt-in synchronization.
Offline-first thinkingDesigned for connection gaps.
A product direction centered on practical access, not a claim that clinical functionality has been deployed or validated.
Core principles
What we're designing for.
The architecture is designed around local inference, structured triage, and patient privacy. Clinical performance and field suitability still require validation.
Clear, actionable information
Use structured symptom questions to support triage recommendations: manage at home, seek local care, or seek emergency care.
Responsible design
Keep data on-device by default; opt-in cloud sync, clinical review, and community health worker feedback are part of the development plan.
Illustrative workflow
From a question to a clearer next step.
A simplified illustration of the intended community health worker workflow.
01
Share context
Answer short, structured questions about symptoms and context.
02
Review information
Receive a triage urgency recommendation grounded in reviewed clinical sources.
03
Consider next steps
Use guidance to consider referral or escalation, with human oversight and validation planned.
Important: PocketSage has a working prototype and is preparing for clinical validation and field pilots. The tool is intended to support triage, not diagnose disease or replace a clinician. Clinical accuracy, safety thresholds, regulatory status, and field effectiveness have not been established.
Technological Architecture
Designed for on-device intelligence.
Our technical approach combines a compressed small language model with structured clinical workflows and a local knowledge library.
Local AI inference
The described architecture uses a 4-bit quantized small language model on consumer Android hardware; inference does not require cloud access.
Clinical grounding
The clinical knowledge strategy references WHO guidance, SNOMED CT, BMJ Best Practice, and PubMedQA. Clinical review and benchmarking are ongoing or planned.
Regional adaptation
Modular LoRA adapters are intended to support disease profiles tailored to regional needs, such as malaria, dengue, and tuberculosis.
Development roadmap
Prototype today. Validation next.
The project is at the pre-deployment / pilot stage. The following are planned steps, not completed milestones.
01
Clinical validation
Benchmark model performance, review triage logic with clinicians, and define safety thresholds, especially for under-triage.
02
Community health worker pilot
Plan a supervised field pilot with approximately 50 CHWs to assess agreement, usability, and device performance, subject to partners and ethics review.
03
Refinement and expansion
Use pilot feedback to improve the tool and consider additional regions or languages only when safety criteria are met.
Stay connected
Interested in the PocketSage vision?
Explore ways to collaborate on offline-first AI triage and responsible clinical validation.
Connect with PocketSage ↗