Mobile Vision
The Vision
A medical-specialised AI agent, in every clinician's pocket.
The phone stops being a tape recorder and becomes the agent — capturing, reasoning and acting on-device. Beyond per-seat licences, value compounds through a combined seats & usage model — more ARR per clinician, and more clinicians.
Business model · land & expand
Seats & usage — grow ARR in both directions at once
Landed · seats Expansion headroom Combined growth
ARR per user Users → Land N seats already landed More ARR same users · usage & actions New ARR more users · seats seats × usage, combined
Pillar 1

Multi-Modal Edge Computing

Every modality — audio, voice, vision — processed on the device. Instant, private, offline-first; the cloud is an optional safety net, never a dependency.
🎙️01

Instant, On-Device Transcription

9:41📶🔋
Live transcript👥
REC00:41
● Layer 1 · ASR✦ Layer 2 · Gemma 4
Speaker 1
Good morning. What brings you in today?
Speaker 2
A sore throat and fever for three days, harder to swallow.
● Live · ASR draft
no cough but the glands feel tender ▏
❚❚ Pause
The two-layer pipeline
Instant draft, then confirmed & labeled
Layer 1 · Streaming ASR~0.3s · live
no cough but the glands feel tender▏
Gemma 4 confirms wording & assigns the speaker
Layer 2 · Gemma 4 · on-device✓ confirmed
Speaker 2: No cough, but the glands feel tender.
Grey = instant ASR draft  ·  solid + labeled = Gemma-confirmed. Both run locally.
🗣️02

Local Speaker Labeling

9:41📶🔋
Live transcript👥
✓ Your voice recognised
Dr. Chen · you
How long has the pain been there?
Patient
About a week now, mostly in the evenings.
Dr. Chen · you
Any fever or chills with it?
Patient
A little — and I've been very tired.
Learns who you are
Your voice → your name, automatically
Enrolled once
🎙 Your voice profile
Heidi learns the clinician's voiceprint, kept on-device.
matched locally, every session
You
Auto-named “Dr. Chen”
No more “Speaker 1” — it knows it's you.
Everyone else
Patient · Speaker 2 · 3…
Other voices separated and colour-coded.
🛡️03

The Backend Accuracy Safety Net

On-device handles everything by default. When a span is uncertain, an opt-in cloud pass double-checks it — and only that snippet ever leaves the phone.
📱
Always · on-device
Instant draft
Private, offline and immediate — the default path for every word.
low-confidence
spans only
☁️
Opt-in · cloud
Verification pass
Re-checks ambiguous terms, drug names and dosages against a larger model.
merge
Result
Confidence-flagged note
Speed by default, assurance where it matters — anything uncertain clearly marked.
📷04

Visual Scribing

9:41📶🔋
Scan & attach
HR78 bpm
BP120/80
SpO₂98%
Temp36.8°C
Objective · OCR
HR 78 · BP 120/80 · SpO₂ 98% · 36.8°C · auto
📎 monitor-vitals.jpg
Original image attached to encounter
Camera as a scribe
OCR the values — keep the proof
Capture
👁 Point at any readout
Monitors, labels, paper forms, wound photos.
OCR
Structured values
Typed into the Objective field, no keying.
Attachment
Original image
Kept on the encounter as evidence.
Pillar 2

Invisible Ambient Workflows

The interface disappears. Control the session by feel, glance and voice — eyes on the patient, and the output lands in two ready-made forms.
🔒05

Lock-screen Controls

9:41📶🔋
9:41
Monday 16 June · 🔒 locked
9:30Jordan Miller · follow-up
9:45Aisha Patel · new
Heidi · compiling note 04:12
▶ Start
✓ Push to EHR
Zero-unlock workflow
Run a whole session without opening the app
▶ Start · ❚❚ Pause · ✓ Push to EHR
Live Activity actions, right on the lock screen.
Glanceable progress
Schedule + note-compile status, always visible.
📳06

Haptic Feedback

9:41📶🔋
Recording · pocket mode
10:00
heartbeat · live & listening
soft pulse · every 5 min elapsed
sharp buzz · audio too noisy
Instantly recognisable signals
Status you feel — no glancing required
Live & listening 5-min time check Audio too noisy
A gentle pulse every 5 minutes keeps the doctor aware of visit length — without ever looking down.
07

Conversational Action Anchors

A natural phrase drops an anchor on the transcript timeline, pinned to the moment. When the visit ends, every action is already drafted and ready for review.
9:41📶🔋
Live transcript👥
REC12:18
12:02 · Patient
The headaches are worse on the left side.
12:11 · Dr. Chen
Let's get you seen by a specialist quickly.
12:18 · voice command
⚓ Action anchor
“Flag for an immediate neurology referral”
✦ Agentic flow · drafting referral…
❚❚ Pause
9:41📶🔋
Visit ended · 12:30
📄
Neurology referral
drafted from 12:18 anchor
💊
Migraine prescription
drafted · awaiting review
✉️
Follow-up email
drafted · awaiting review
Review & approve all (3)
Spoken → anchored → drafted
The conversation does the paperwork
Each anchor stays linked to its moment. By the time the patient leaves, the documents are written — you just review and approve.
08

The Dual-Stream Output

One conversation, two ready-made outputs — clinical depth for the chart, plain language for the patient.
9:41📶🔋
Visit output
🩺 Clinician
🧑 Patient
Stream A · for doctors
Assessment
Acute viral pharyngitis
Plan
Symptomatic management; safety-net advice
Review in 7 days if persisting
🔖 NICE CKS · sore throat
Coding
J02.9 · suggested
9:41📶🔋
Visit output
🩺 Clinician
🧑 Patient
Stream B · for Aisha
💧
Rest & drink water
plenty of fluids
💊
Paracetamol if sore
2 tablets, up to 4× a day
📅
Come back in 1 week
if you're not better
Scan to take it home
Pillar 3

Agentic Flows

Heidi doesn't just transcribe — it acts. Each action is drafted from the conversation, then reviewed and approved by you before anything happens.
09

Agentic Actions, by Example

Drafted
From the conversation
👀
You review
Edit anything
✍️
You approve
Sign it off
It runs
Sent & filed
🗂️
Chart template
Picks the right note layout for the visit.
Suggested template
✓ Respiratory consult best match
General consult
Mental health review
Change
Use
✉️
Follow-up email
Drafts the patient note, ready to send.
Draft email
To
aisha.patel@email.com
Subject
Your visit today & next steps
Hi Aisha, thanks for coming in. As discussed, rest and fluids…
Edit
Send
💊
Prescription draft
Accurate Rx pulled from the conversation.
Prescription · draft
Amoxicillin 500 mg
1 capsule · three times daily · 7 days
⚠ Caveat: penicillin allergy — none on file
Edit
✓ Approve
🔖
Citation-backed evidence
Links the plan to a guideline source.
Plan · annotated
Symptomatic management; safety-net advice.
🔖 NICE CKS · sore throat ↗
Source attached to the note for audit.
View source
Keep
⚠️
Medical caveats
Flags interactions & contraindications.
⚠ Safety flag
Ibuprofen + warfarin
Increased bleeding risk — consider paracetamol instead.
High severity · review before approving
Dismiss
Swap drug
⤴️
Export to systems
Pushes the approved output downstream.
After approval
Note → Best Practice EHR
Script → pharmacy
Plan → patient portal
⤴️ Export all
Pillar 4

All-in-one Workstation

For the desk, not just the pocket — a shared, always-on Heidi station. Walk up and it knows you; the screen never sleeps, with live transcription and agentic actions woven in.
🖥️10

The Heidi Station

Heidi · Working● listening
Dr. Chen  Let's check your blood pressure.
Patient  Okay, sure.
✦ Vitals captured → Objective
Dr. Chen  I'll arrange a referral today.
✦ Referral drafted · ready to review
Scan to take
your notes
3 actions ready
📟
Customised Heidi tablet
Clinical-grade Android tablet, leaning on a soft fabric dock.
🔌
Dock charges all three
Cradles the tablet and two Heidi remotes at the front, charging together.
📡
Always-on live transcription
The screen never sleeps — the conversation appears as it happens.
🔗
Remote auto-sync · no pairing
Drop a Heidi remote on the dock and it syncs to the tablet instantly — no manual pairing.
🙂
Face recognition · auto-profile
A glance loads the right clinician's templates, macros and voiceprint — on a shared station.
On-screen QR takeaway
The patient scans the screen to take their notes and plan home.
Team Enablement
Make the team AI-native.
Give every engineer a shared harness, self-running loops and a clear fluency ladder — so AI compounds across the team, not just the individual.
🧰01

The AI Harness

Shared scaffolding that makes building with AI fast and safe — quality is automatic, not heroic.
🛠️
Build
you + agent, in flow
Harness verifies
tests · design system
🎬
Auto-report
shots & recordings
🔁
Feedback
back into the harness
🧪
Full test harness
Unit, UI, integration & E2E — every change verified automatically.
🎨
Design system
Consistent, reusable UI building blocks.
🖥️
Mockup local server
Spin up and preview any screen instantly.
🤝
Team subagents & skills
Shared agents and skills anyone can invoke.
🎥
Auto screenshots & recordings
Self-documenting reports and demos.
📚
Compliance knowledge base
Rules & guardrails the agent must follow.
♻️02

Autonomous AI Loops

Agents that run themselves — triggered by events and schedules, doing the recurring work in the background.
Triggers
Event-triggered
🌙 Nightly jobs
🔁 CI/CD pipeline
🔍 Regular scans
🤖
Agent loop ↺
Outputs
🐞 Bug reports
📊 Analytics
🔀 PRs for review
🎥 Demos for discussion
📚 Knowledge-base updates
📈03

AI Fluency Ladder

A shared language for how people work with AI — used in hiring and performance reviews.
0
Manual
Doesn't use AI, or only for searching knowledge.
1
Assisted
Turn-based agents for simple code changes.
2
Orchestrating
Workflows and multiple subagents for bigger-scope tasks.
3
Harness-builder
Builds team harnesses; defines auto-triggered systems and ways of working.
4
Paradigm-setter
Brings and proves new agent paradigms at org or company level.
🐶04

Dogfooding

The team is the first user — ideas become installable builds and get hardened in-house before customers ever see them.
Spark
💡 Hack ideas
Anyone prototypes a new agent or workflow.
Ship internally
📦 Internal distribution
Promising builds reach the team first — easy to install.
Harden
🧪 Alpha test ground
Real internal usage hardens it before release.
↺ What we learn flows straight back into the harness and the loops.
Operational Excellence
Healthy by default, fixed before users notice.
Full-stack monitoring, an agent that closes the loop on issues, and a tight line from customer signal to fix.
📡01

Full Monitoring

From system vitals to business funnels — one live view of what's healthy, watched against SLOs with error budgets.
Operations · live all systems healthy
99.95%
Uptime
✓ within budget
1.2s
p95 latency
✓ on target
0.03%
Error rate
✓ healthy
Record → note generated ok · 1.4s
Approve → push to EHR ok · 0.9s
Patient handout / QR slow · 3.1s
Encounter100%
Note drafted96%
Approved88%
Exported84%
What we watch
Vitals up to funnels
Golden signals
⚡ Vitals
Latency, traffic, errors, saturation.
Does it work?
🧭 Critical journey health
The flows that must never break.
Does it convert?
📊 Critical funnels
Where users drop off, step by step.
Always-on
🔔 Monitors & alerts
SLO-based, actionable, owned — no noise.
🛠️02

Auto-Remediation

Monitoring doesn't just page someone — an agent triages the issue, drafts the fix and opens a PR for review.
🔔
Detect
alert fires
🔎
Triage
agent finds root cause
🔀
Draft PR fix
with tests
👀
Human review
approve & merge
Verify
alert clears
Example · auto-drafted fix
A PR, opened for review
🔀 Pull request · agent
Fix: handout QR timeout under load
Root cause: blocking call on the render path.
+24 −8 · 2 files · tests added
✓ CI green · p95 back under 1.5s in canary
View diff
✓ Approve
Example · auto ops report
Generated weekly, for you
🗂 Operational excellence · this week
Uptime 99.95% · within error budget
3 issues auto-fixed & merged
p95 latency −18%
⚠ 1 funnel regression flagged for review
Open full report
💬03

Voice of the Customer

Every signal — reviews, support, sentiment — clustered and groomed into a prioritised backlog, not a pile of tickets.
Signals in
App-store reviews
🎫 Support tickets
📈 NPS & sentiment
💡 In-app feedback
🧹
Groom & cluster ↺
Backlog out
🐞 Bugs · deduped
Feature requests · ranked
📊 Themes & sentiment trend
🗺️ Roadmap input
⚙️04

Operating Principles

The common ground rules that keep all of the above honest.
🎯
SLOs & error budgets
Set reliability targets; spend the budget on shipping speed.
🔭
Observability
Metrics, logs and traces — ask why, not just what.
🔔
Actionable alerts
Every page is real, urgent and owned. Kill the noise.
🚨
Incident response
Clear severities, on-call and comms when things break.
📝
Blameless postmortems
Fix the system, not the person. Every incident teaches.
🚦
Progressive rollouts
Canary, monitor, then ramp — with fast rollback.
Ways of Working
One person, a team of agents.
A repeatable AI-driven workflow with human checkpoints — plan, build in parallel, and review at the right altitude.
🧭01

The AI-Driven Workflow

Every feature runs the same path — planning with human checkpoints, parallel agent build, then a high-level human review.
① Plan
🗂️ JIRA epic
📄 PRD 👤 review
🏗️ System design 👤 review
📋 Implementation plan
Acceptance & test plan
② Build
🧩 Break down tasks
🤖 Agents build in parallel
🌿 on a feature branch
③ Review
👤 Human review high-level
🎯 Use-case check
🔀 Merge & ship
🤖02

Parallel Agents

On a feature branch, specialised subagents work at once — one builds, one verifies, one reviews — converging into a single human checkpoint.
Isolated
🌿 Feature branch
one per feature.
fan out
🛠️ Implementation
writes the code.
🧪 Verification
runs tests & checks behaviour.
🔍 Review
critiques the diff.
converge
Gate
👤 Human review
high-level + use case, then merge.
🌗03

The Day / Night Rhythm

Humans and agents work different shifts — daytime for judgement, overnight for the long haul.
☀️
Daytime · human time

Communicate & verify

  • 💬 Communication & alignment
  • Human verification & decisions
  • 🎬 Demos and reviews
  • 🧭 Steering the agents
🌙
Overnight · agent time

Long-horizon work

  • 🖥️ Long tasks on a VM
  • 🏗️ Big refactors & batch work
  • 🧪 Exploration & spikes
  • 🌅 Ready for you in the morning
🧑‍✈️04

Operating Habits

The habits that make a one-person team work.
📓
agents.md
A living local knowledge base — agents read it and auto-update its rules.
🎬
Demo for everything
Every change ships with a demo — the unit of communication.
🧑‍✈️
One-person team
One person orchestrates a fleet of agents, end to end.
Technical Direction
Built for the edge.
On-device models, agentic interfaces and an architecture that keeps PHI on the phone — riding the 2026 edge-AI wave.
🧱01

The On-Device AI Stack

Everything from the runtime up runs on the device — private, instant and offline by default.
🧩
App APIs
Structured output · tool-calling · prefix caching · streaming ASR
build
🎛️
Adaptation
Fine-tuned local models · per-task LoRA adapters
specialise
🧠
Models
Gemma 4 E4B (multimodal, low-RAM PLE) + Gemini Nano 4 via Android Prompt API
reason
⚙️
Runtime
LiteRT-LM · GPU + NPU via AICore · memory-mapped · zero network calls
run
🪄02

Agentic & Generative UI

The interface assembles itself around intent — the agent picks the right tools and data for the moment and renders just what's needed.
Signal
🎯 Intent
what the clinician needs now.
Agent
🧠 Reason & discover
finds tools (MCP), returns structured output.
Render
🪄 Generated UI
only the controls & data for this moment.
🧭
Intent-based navigation
Surface the task, not a menu tree.
🧱
Generative components
Buttons, charts and layouts built on the fly.
🔭
Tool discovery · MCP
Agents discover tools, not just call them.
🛰️03

BFF & Control

A backend-for-frontend wraps the shared core backend — the mobile app gets its own control plane (flags, model routing, policy) while PHI stays on the device.
📱
Edge
Device
Gemma 4 / Nano run here; PHI never leaves.
flags · routing
policy · telemetry
🛰️
Control plane
BFF
Backend-for-frontend — tailored control for the mobile app, no app release.
one core,
many clients
🗄️
Core
Shared backend
The real backend shared with web & other frontends — EHR, sync, data.
📝04

Auto Feature Request

Demand becomes structured feature requests automatically — from the sales floor and from real usage.
Presale
🤝 Sales signals
Lost-deal reasons, prospect asks, RFP gaps.
Users
📲 Usage signals
Friction points & in-app requests.
Agent
🤖 Auto-draft request
Structured, deduped and sized.
Outcome
🗺️ Prioritised roadmap
Evidence-backed, ready to plan.
🔭05

Horizon Bets

The edge-AI moves we're backing next — each one unlocks a high-priority request.
🧩
Structured Output API
Typed outputs that drive UI and tool-calls reliably (ML Kit GenAI, 2026).
↳ reliable agentic actions
🛠️
On-device tool-calling
Local function-calling so the agent can safely act, not just suggest.
↳ prescriptions · referrals · export
Live agentic actions
Actions drafted inline as the conversation unfolds, not after.
↳ actions in the transcript
🧬
Private on-device personalisation
Per-clinician LoRA that adapts locally, privately.
↳ per-clinician profiles
🔋
Power-aware NPU inference
INT4 quantization + NPU offload via AICore — the biggest lever on battery.
↳ all-day on-device use