HealthIA ONE

Your health never starts over — a patient-owned AI agent that turns evidence into durable, consent-aware missions and real-world action.

Comment

HealthIA ONE

Your health never starts over.

Most health AI waits for the next prompt. HealthIA ONE is designed to keep safe health work moving because the patient’s authorized longitudinal state changed — not only because the patient remembered to ask another question.

HealthIA ONE is a patient-owned, event-driven health continuity agent. It turns fragmented health signals, treatment context, clinical evidence and unfinished work into durable missions that can survive the conversation, react to new evidence, act through bounded Google services, stop at human-authority boundaries and close only when verifiable evidence exists.

Chat is one interface into HealthIA. It is not the source of truth.

Official live product demo

YouTube (~3:17): https://youtu.be/44LfVn9pPdU

The current V5 demo stays inside the real HealthIA ONE product running on Google Cloud. All patient data shown is synthetic.

1. Treatment-aware Result Guardian — no new chat prompt

The synthetic Patient Twin already contains a clinician-confirmed Losartan 50 mg daily treatment.

A renal-function result containing creatinine is uploaded through the real Results workspace. HealthIA evaluates the new evidence against longitudinal treatment context and detects that relevant potassium evidence is still missing.

Without a new patient chat prompt, HealthIA opens a durable treatment-aware mission so that unfinished monitoring work cannot disappear when the conversation ends.

When potassium evidence later arrives through the real product, HealthIA links it to the existing mission and closes that same mission from durable evidence.

HealthIA does not diagnose kidney disease and does not change the medication.

2. Appointment Guardian — the Twin prepares the visit

A family-medicine follow-up requires recent results, an active medication list and insurance.

HealthIA verifies what the longitudinal record already contains. Results and medication context are present; insurance evidence is missing, so HealthIA creates an appointment-preparation mission.

When the patient uploads the insurance document through the real Documents workspace, HealthIA re-checks the requirements and closes the same mission.

3. Post-Visit Guardian — continuity survives the encounter

When the appointment becomes completed but HealthIA cannot verify an attributable consultation note or discharge document, Post-Visit Guardian opens a continuity mission instead of inventing what happened.

When the consultation note later arrives, HealthIA correlates the evidence to the completed encounter and closes the same mission.

4. Live Gemini 3.5 Flash + Google ADK adaptive interview

Inside the real HealthIA Chat, Gemini 3.5 Flash generates a bounded, case-specific adaptive interview.

The recording gate requires the live response to report gemini_dynamic and exactly five case-specific questions before the scene is accepted. Google ADK coordinates the clinical capabilities while the durable Patient Twin remains the continuity layer underneath the conversation.

5. Real autonomous external follow-up

This is the strongest Taskmaster proof in the current film:

no new chat prompt
→ overdue blood-pressure mission
→ Eventarc wakes a private worker
→ real Gmail message sent
→ controlled patient reply in the same thread
→ Gmail users.watch
→ authenticated Pub/Sub
→ VitalRecord 128/80 with source_type=patient_email_reply
→ the same durable mission becomes COMPLETED

The outside-world result is not accepted because a model says it happened. Completion comes from durable external evidence returned through the real event path.

Why this is agentic rather than just conversational

The core HealthIA loop is:

Signal / new evidence
→ Patient Twin
→ Guardian evaluation
→ durable mission
→ safe action or evidence wait
→ real connector result / durable evidence
→ verified closure
→ updated Twin

The patient does not need to repeatedly reconstruct context in chat. HealthIA can notice unfinished work because authorized state changed.

Durable Patient Twin instead of prompt memory

Firestore stores patient-scoped missions, treatment context, consent, evidence references, appointments, connector outcomes, device signals and longitudinal state.

The Patient Twin is derived from canonical patient state rather than being a second independent database or a long prompt. Missions and evidence can survive logout/login and Cloud Run process replacement.

Evidence before interpretation

For supported synthetic clinical PDFs, images and structured result files, original evidence is preserved first in private Google Cloud Storage. Gemini performs bounded extraction where appropriate, structured results are persisted with provenance and the Patient Twin links derived state back to original evidence.

If reliable extraction cannot be established, HealthIA fails closed instead of fabricating a finding.

Mission-scoped consent and human authority

HealthIA separates three decision modes:

  1. AI reasoning for interpretation, multimodal understanding and adaptive questioning.
  2. Deterministic logic for exact state transitions, idempotency, policy and safety invariants.
  3. Human authority for consent and clinically sensitive decisions.

HealthIA can create durable work before a sensitive authorization exists, but it does not execute a consent-gated connector merely because a mission exists.

It does not autonomously diagnose, prescribe, start/stop/change medication or replace professional or emergency evaluation.

ONE SAFETY — authorization is not execution evidence

Protected real-world actions cross a narrow audited chain:

Sense / Request
→ Reason
→ Authorize
→ ONE SAFETY / deterministic Safety Kernel
→ one-use HealthActionTicket
→ real connector
→ durable receipt
→ OpenTelemetry / Google Cloud Trace
→ mission outcome

A model-generated success message cannot make an external mutation true. A connector outcome and durable receipt are required before completion can be projected.

The repository also preserves a promoted ONE SAFETY proof that correlates a protected Google action across Cloud Trace, HealthActionTicket, connector receipt and completed mission outcome.

Prompt-injection boundary

HealthIA uses both Google Model Armor and a local fail-closed ingress policy.

The adversarial proof requires all of these simultaneously:

  • Google Model Armor detects the controlled jailbreak probe;
  • HTTP 400 at prompt ingress;
  • model_called == false;
  • zero new HealthActionTickets;
  • zero patient-state mutation.

Technologies used

  • Gemini 3.5 Flash on Vertex AI / Google GenAI — reasoning, adaptive clinical questioning and multimodal interpretation
  • Google Agent Development Kit (ADK) — agent framework and capability orchestration
  • Google GenAI SDK
  • Cloud Run — application and bounded private workers
  • Firestore — canonical patient-scoped durable state
  • Private Google Cloud Storage — original clinical evidence
  • Eventarc — event-driven worker activation
  • Gmail API + Gmail users.watch + Pub/Sub — autonomous external follow-up loop
  • Google Model Armor — prompt-injection / jailbreak boundary
  • Secret Manager + ADC / dedicated service identities
  • Google Places / Maps Platform
  • Google Calendar + Google Tasks
  • OpenTelemetry + Google Cloud Trace
  • Android / Health Connect bridge contracts

Demo integrity

The V5 judging film is composed from a real continuous product recording. Long infrastructure wait periods caused by Cloud Run/Eventarc/Gmail/Pub/Sub latency were removed to fit the judging window.

No Guardian transition, Gemini answer, Gmail send, patient reply, 128/80 VitalRecord or mission completion was recreated or substituted. The functional truth gate had to pass before the film could be accepted.

Demo and production-readiness links

The README includes a from-scratch deterministic setup that can be run with zero Google AI request budget, plus technical verification commands.

Findings and learnings

  • Durable state matters more than longer prompts.
  • The Twin should create work, not merely summarize history.
  • More AI is not always better AI: exact state transitions and safety invariants are safer when deterministic.
  • Authorization and execution evidence must remain separate.
  • Useful autonomy needs human boundaries, not the absence of them.
  • Event-driven autonomy is easier to audit than an unconstrained permanent agent loop.

Bonus evidence

The bonus paths remain isolated from the frozen final runtime.

Clinical truth boundary

HealthIA ONE is a synthetic hackathon prototype and patient continuity system. It is not a physician, emergency service, autonomous prescription engine, regulated medical device or clinical-effectiveness study.

All hackathon patient data is synthetic. Green software and Cloud proofs establish tested software behavior; they do not establish medical efficacy, regulatory approval or universal security certification.

The one sentence to remember

HealthIA carries unfinished patient work forward, notices when the authorized Twin changes, does every safe step it can prove, and stops exactly where human authority begins.

Built With

  • android-health-connect
  • cloud-run
  • fastapi
  • firebase-cloud-messaging
  • firestore
  • gemini-3.5-flash
  • gmail-api
  • google-agent-development-kit-(adk)
  • google-calendar-api
  • google-cloud
  • google-cloud-text-to-speech
  • google-genai-sdk
  • google-places
  • google-tasks-api
  • playwright
  • pub/sub
  • secret-manager
  • vertex-ai
Share this project:

Updates