Digital health · HIPAA-aligned multi-agent AI
LifeLine
A virtual mental-health treatment platform with chatbot, medical-scribe, and risk-analytics agents on Amazon Bedrock, Chime SDK video, and a REB-approved safety evaluation.
- Role
- Senior Software Engineer, Cloud & AI · led development, cloud infrastructure and solution architecture end to end
- Organization
- Virtual Health Hub
- Timeframe
- 2024 – 2026
- Amazon Bedrock Agents
- Amazon Chime SDK
- Amazon Transcribe
- Amazon Polly
- ECS Fargate
- Amazon Cognito
- Amazon Kinesis
- React Native (Expo)
- Next.js
- Vue.js/Quasar
- Node.js
- PostgreSQL/Prisma
Context
LifeLine is a virtual mental-health treatment platform built at Virtual Health Hub. It serves Northern Saskatchewan, including Indigenous and remote communities, through a mobile app on the App Store and Google Play: the Pimâtisiwin chatbot, virtual consultations, scheduling, medication management, mood journaling, and questionnaires, with clinician portals on the care-team side. The constraints were specific. Protected health information had to be handled inside a HIPAA-aligned architecture: PHI-handling services in a HIPAA-aligned VPC, with AWS managed services for AI, media, and voice. Generative AI talking to mental-health patients had to be evaluated for safety, not assumed safe. And consultation video had to run inside a React Native (Expo) app through AWS Chime SDK, which meant writing native Swift and Kotlin bridges. Mahmood Ul Islam managed development, cloud infrastructure, and solution architecture end-to-end.
The build
Separate Bedrock agents, not one prompt.
The AI is split into a patient-facing chatbot agent, a medical-scribe agent for documentation, and a risk-analytics agent, all built on Amazon Bedrock Agents. Each agent has a narrow job. The system uses RAG so answers come from retrieved content rather than model memory alone.
Voice through Amazon Transcribe and Amazon Polly.
Speech-to-text and text-to-speech sit in front of the chatbot so a patient can speak and listen instead of typing.
Amazon Chime SDK video with ambient-listening summarization.
Virtual consultations run on Chime SDK, and ambient listening summarizes the session's conversation.
A native Chime bridge for React Native, released as open source.
The Expo app needed the native Chime SDK on both platforms, so Swift (iOS) and Kotlin (Android) Expo modules were written and packaged as a TurboModule/Fabric library: react-native-chime-sdk, Apache-2.0. It covers multi-party video, audio, content share, blur, and data messages, and the production app uses the published package. TurboModule and Fabric are React Native's new-architecture APIs.
HIPAA-aligned VPC on ECS Fargate.
Backend services run on ECS Fargate inside a HIPAA-aligned VPC, with PostgreSQL (through Prisma) as the data layer, Amazon Cognito for identity and access, and Amazon Kinesis for streaming data. Alignment means the network, identity, and service choices were made with HIPAA's requirements in view. Compliance is a legal and audit determination, not something an engineer can declare.
Biometrics and clinician portals.
Apple HealthKit and Health Connect integration auto-syncs 30+ biometrics to the care team. Clinician portals were delivered in Vue.js/Quasar and Next.js. Appointment booking and medication management were built with analytics.
A REB-approved safety evaluation of the agentic AI.
Rather than assert the agents were safe, the founder was first author of a safety evaluation approved by a research ethics board: 4,500+ scored interactions across two datasets, validated against expert clinical review. The manuscript was prepared for Mayo Clinic Proceedings: Digital Health.
Outcome
4,500+
scored interactions in the REB-approved safety evaluation
30+
biometrics auto-syncing from HealthKit and Health Connect
Apache-2.0
react-native-chime-sdk, open source and used in production
iOS + Android
live on the App Store and Google Play
- LifeLine is in production: available on the App Store and Google Play, serving Northern Saskatchewan, including Indigenous and remote communities.
- The platform's agentic AI went through a REB-approved safety evaluation: 4,500+ scored interactions across two datasets, validated against expert clinical review, with a manuscript prepared for Mayo Clinic Proceedings: Digital Health (Mahmood Ul Islam is first author).
- 30+ biometrics auto-sync from Apple HealthKit and Health Connect to the care team.
- react-native-chime-sdk is published under Apache-2.0 and used in the production app.
Engineering notes
Why separate agents
One large prompt is simpler to ship and harder to reason about. Splitting chatbot, documentation, and risk analytics into separate Bedrock agents costs more orchestration, but each agent gets a narrow scope and can be evaluated on its own inputs. In a mental-health setting that separation is the point: the agent that talks to a patient and the agent that analyses risk should not share one set of instructions. RAG keeps answers tied to retrieved content rather than model memory alone.
HIPAA-aligned, stated carefully
PHI-handling services run on ECS Fargate inside a VPC designed for HIPAA alignment, identity and access go through Amazon Cognito, and the AI, media, voice, and streaming paths use AWS managed services (Bedrock, Chime SDK, Transcribe, Polly, Kinesis). We say "aligned" on purpose. HIPAA compliance is a legal and audit determination, not a badge an engineer can award. "Aligned" is what engineering can deliver, and it is the word we will use with you too.
Evaluate like a study, not a demo
The safety evaluation was approved by a research ethics board, and its scores were validated against expert clinical review rather than internal judgment alone. 4,500+ interactions across two datasets is enough to see patterns, not anecdotes. On the mobile side, the Chime bridge was built as a TurboModule/Fabric library and released under Apache-2.0; the production app depends on the public package, which keeps the API honest and the versioning disciplined.
Relevant if
Multi-agent AI on regulated data. Separate Amazon Bedrock agents with narrow jobs, the way LifeLine splits chatbot, medical scribe, and risk analytics, with RAG, inside a HIPAA-aligned VPC.
Video and voice inside a mobile app, with AI on top. Chime SDK video with ambient-listening summarization, Transcribe and Polly voice, and native Swift and Kotlin bridges for React Native (Expo), the work published as react-native-chime-sdk (Apache-2.0).
A safety evaluation before an LLM feature reaches patients or customers. Study design, scored interactions, LLM-as-judge evaluation, and a protocol your clinical experts can validate against. Mahmood Ul Islam is first author of LifeLine's REB-approved safety evaluation (4,500+ scored interactions across two datasets).