Field
Turning the voices of a complex on-scene interactions into structured medical notes - all offline.
Why this Project?
AI that structures messy input into compliant output — the same pattern as CPQ.
Field takes unstructured voice and extracts structured, validated data in real time. That's the core challenge Roadrunner is solving: take a complex, ambiguous input (a deal) and produce a compliant, structured output (a quote). At an early-stage company co-developing with design partners, the interface IS the product. This project proves I can design AI-native flows from zero — where the model does the heavy lifting and the user confirms.
Framing
Search and rescue teams still use paper to document patient care. They write in the rain, in the dark, wearing gloves. Important details get lost. When they hand off a patient to a hospital, the notes are often incomplete or hard to read.
Existing software was built for ambulances with Wi-Fi and keyboards. Nothing works for the backcountry — where there's no cell service, no keyboard, and no time to type.
Personas
How It works
The responder talks to the patient. The app listens, logs all medical details, and creates a patient record and SOAP note automatically. The record travels with the patient in a tap.
Core Experience
01
Dictation
Keep treating, the note writes itself
The provider talks to the patient. The app listens, identifies and organizes into compliant medical data in real time — all without touching the screen.
Jobs To Be Done
As a responder, I can...
02
Incident Report
See a SOAP note as you work
A SOAP note, the standard format of capturing patient information, builds itself as the encounter unfolds. Completion counters show progress per section. Stoplight warnings flag values that need confirmation before handoff.
Jobs To Be Done
As a responder, I need...
03
Manual Entry
Always one hand on the patient
Sometimes, data needs to be entered by hand. Capture vitals with cold hands, bad light, divided attention. No keyboard, no precision tapping.
Jobs To Be Done
As a responder, I need...
04
Offline Care Transfer
Continuity of care in a single tap
When the patient is transferred, the critical information goes with them — even offline. An anonymized clinical summary can be sent via Apple Wallet, satellite, or SMS. The receiving team gets the whole case — not a phone call from memory.
Jobs To Be Done
As a responder, I need...
Process
01
Design Decisions
Why is voice the right model?
Your hands are on the patient, not a keyboard. Field listens to the conversation and writes the documentation. We fine-tuned speech recognition for medical terminology and outdoor noise so it works where it's actually needed.
Why allow manual entry?
Even with our fine-tuned models, AI misses things — a vital gets taken silently, some fields are never spoken aloud. Field provides glove-friendly manual input for every gap. Type into the narrative and structured fields auto-populate below.
How do you trust AI output?
Every AI-extracted value shows where it came from. Grey means confident, amber means check it, red means missing. Each value traces to the exact moment in the transcript. The provider always knows what to trust.
Why not let AI make clinical decisions?
When it comes to suggesting what to do with a patient, Field shows recommendations but never pre-selects. The provider makes every clinical call. This keeps us on the right side of FDA regulation and keeps humans in charge of decisions where terrain, weather, and context matter more than data.
Why does it need to work offline?
No cell service on a mountainside. Everything runs on-device. When connectivity returns, Field exports a standards-compliant file into the reporting systems agencies already use. It doesn't replace existing software — it replaces the paper and memory that come before it.
02
Design Systems
BACKGROUND
TEXT
SEMANTIC
SPEAKER IDENTITY
BORDER
SPACING & GRID
DESIGN PRINCIPLES
INFORMATION
Progressive disclosure
Collapsed → expanded. Summaries → detail. Don't overwhelm.
Glanceable
Color dots, coded cells, badge counts. Triage at speed.
INTERACTION
Thumb-reach
72px minimum touch targets. Bottom-pinned actions. One-handed.
Selection = border
#EEF8FE border on bg.elevated. Consistent across all controls.
TYPOGRAPHY
Inter, always
Single typeface. Weight + size = hierarchy. Monospace for numbers only.
SURFACE & COLOR
Darkness-first
Dark surfaces for outdoor and low-light. #EEF8FE on black.
Layered surfaces
App → card → elevated → input. Each step lighter. No shadows.
Semantic color only
Red = critical. Amber = warning. Green = success. Blue = info. No decoration.
ACCESSIBILITY
Color is never the only signal
Every status uses color + text + position. Stoplight vitals pair background tint with numeric value. Badges show count and label. Nothing requires color vision alone to interpret.
Impact
FAQ
Is this available for professional EMS?
Not yet, but it's coming. The same offline-first foundation, voice capture, and compliant data model is being extended to write directly into the ePCR — the electronic patient care report that ambulance crews and 911 call centers rely on.
Does it work offline?
Yes. The app downloads two AI models and runs them entirely on the device, so it behaves the same in town, deep in a canyon, or 30 miles from the nearest cell tower.
Does the dictation actually work in the field?
It does. The dictation model was fine-tuned on more than 5,000 clinical recordings chosen for high wind, rain, helicopter landings, patients in pain, and overlapping voices.
Is it compliant with privacy and data standards?
Yes. Field is built to HIPAA, NEMSIS, FHIR R4, and USCDI v3, so records can move cleanly into hospital systems and state EMS registries.