Technology
Enterprise-grade architecture built on modern, verifiable technologies for reliability, security, and interoperability.
On-Device AI Inference
Unlike cloud-dependent EMRs, ChartLite executes entirely locally. sherpa-onnx powered ASR with 7 model tiers runs fully offline, including medASR for clinical vocabulary accuracy. Quantized Qwen 3.5 runs through llama.cpp for clinical extraction. An on-device RAG pipeline retrieves only the most relevant diagnoses and drugs per encounter.
Zero-Trust Security
Data at rest is protected by a SQLCipher encrypted database. Offline synchronization uses v4 AES-256-GCM encrypted binary payloads (92 bytes → 120 bytes encrypted → 160 Base64 chars = 1 SMS) transmitted over standard SMS networks. The v4 format carries expanded vitals, 5 chronic conditions, immunization history, growth z-scores, and clinical status flags (HIV, TB, pregnancy). Key derivation relies on PBKDF2 with 100,000 iterations. Play Integrity device attestation verifies app authenticity before any cloud API access.
Global Interoperability
FHIR R4 bundles are generated on-demand for external systems including DHIS2 and OpenMRS. Automated ICD-10 to CPT/HCPCS mappings power insurance claims. The architecture is ready for national health information exchanges across 3 active + 5 planned countries.
100% offline, scales as you grow
A tiered architecture that starts with zero connectivity requirements and progressively unlocks capabilities as infrastructure improves.
- Fully functional without internet on standard Android devices — every feature from voice capture to structured encounter works offline out of the box.
- Battery-aware inference queues and batches transcripts, loading the model once for multiple patients instead of per-encounter, preserving charge through a full clinic day.
- When connectivity is available, secure peer-to-peer sync over Bluetooth or WiFi enables multi-device collaboration — both within a facility and across facilities via patient-scoped transfers with user confirmation. Optional cloud AI and enterprise reporting unlock additional capabilities without replacing the local pipeline.
- Adapts from a single-clinician rural healthcare outpost to a fast-paced, multi-provider urban practice — same app, same data model, no reconfiguration required.
Tech stack
Every component is open-source, auditable, and selected for reliability at the edge.
| Layer | Technology | Purpose |
|---|---|---|
| Platform | Kotlin + Jetpack Compose | Native Android UI with declarative components |
| Database | Room + SQLCipher | Encrypted FHIR-native schema, zero-setup local storage |
| ASR | sherpa-onnx (7 model tiers) | Multi-language speech recognition, fully offline, including medASR for clinical use |
| NLP | llama.cpp (quantized Qwen 3.5) + 5 cloud fallbacks | 6-strategy extraction: Gemini Nano → Qwen on-device → Claude → OpenAI → Gemini → Regex. Flash attention, Q8 KV cache |
| Encryption | AES-256-GCM + PBKDF2 | End-to-end data protection at rest and in transit |
| Sync | Bluetooth / WiFi P2P + SMS v4 fallback | Peer-to-peer data exchange with cross-facility patient-scoped sync; v4 SMS packs 92 bytes (34B encounter + 38B health history + 19B free text + 1B CRC) into a single 160-char message |
| Billing | ICD-10 → CPT/HCPCS, E/M Levels, SOAP | Insurance claims engine with automated coding and SOAP note generation |
| Standards | FHIR R4, ICD-10, CPT/HCPCS, DHIS2 | Interoperability with national health systems and insurers |
| Security | Play Integrity API | Device attestation for cloud API authentication |
| License | Apache 2.0 | Always free, always auditable, no vendor lock-in |
Go deeper
Explore the technical details behind each layer of the architecture.
On-device intelligence
How sherpa-onnx ASR, Qwen 3.5, and retrieval-augmented extraction work together on a $100 phone.
Read more →Sync & data strategy
Three-tier connectivity from fully offline through Bluetooth peer sync to cloud-connected enterprise reporting.
Read more →Security architecture
SQLCipher encryption, AES-256-GCM v4 SMS payloads with immunization and clinical status flags, and zero-trust design across every data path.
Read more →