← David Randolph SIDE PROJECT 01 / 04
Motion graphic · 24-second loop · voice is computer-generated · illustrative

CONCEPT · AI · HEALTHCARE · MOBILE

Designing an AI note-writer that nurses actually trust

Home-care nurses often write visit notes at night, from memory. Tend is my concept for an AI note-writer that turns the voice memos they already record into a draft — designed so nurses stay the author, not the proofreader.

MY ROLE
Self-initiated, solo — research, UX, UI, prototyping
PLATFORMS
Mobile (iOS, Android)
TIMELINE
Self-initiated, 2026
METHODS
Secondary research, design principles, coded prototype

A self-initiated concept. Tend is an invented product name; screens are illustrative.

Same day
Goal: notes finished between visits, not at midnight
Every line
Goal: traced to a memo clip or photo
100%
Goal: drafts reviewed section by section before signing
0
Guardrail: AI-invented values in a signed note

Concept goals, not measured results.

01 — THE PROBLEM

"Add AI" isn't a brief

Home-care nurses visit patients at home and document each visit in a long structured form. Notes are often written at night, from memory — slow for nurses and risky for patients. Meanwhile, many healthcare teams are being asked for an "AI strategy" with no specific problem to point it at.

My first job was to narrow it: find the one moment where AI would save the most time with the least risk.

02 — THE INSIGHT

The note already exists — as a voice memo

The insight this concept is built on: nurses often record quick voice memos between visits so they don't forget details, then retype everything later. The raw material for the note already exists. It's just trapped in the wrong format.

The risk is trust. A visit note is a legal record, and the nurse signs their name under it. Any AI that feels like a black box would be ignored or, worse, trusted blindly.

03 — DESIGN PRINCIPLES

Three rules for trustworthy AI

Every line has a source
Tap any sentence to hear the memo clip or see the photo it came from.
Gaps, not guesses
If something required is missing, the AI flags it — it never fills it in.
The nurse signs, always
Drafts can't be submitted without a review step. Friction, on purpose.
04 — PROTOTYPING

Designing the draft to be checked, not skimmed

A long, confident AI paragraph invites skimming, and skimming is how errors get signed. So instead of a long-form draft, the prototype breaks the note into the form's own sections, with source links and highlighted values.

Next, I'd test it Wizard-of-Oz style: nurses send real memos, a clinician writes the "AI" draft by hand, and we count how many planted errors reviewers catch in a long-form draft versus the sectioned one.

Recording · 0:42
Talk like you already do between visits.
Capture fits the habit nurses already had.
WOUND ASSESSMENT · DRAFT
Left heel wound measured 2.1 cm, down from 2.6 cm. Edges pink, no drainage.
Memo 0:12–0:24Photo 10:14
PAIN
Patient reports 3/10, improved.
Drafts follow the form's sections; every value links to its source.
Review before signing
Missing: medication changes. Add or confirm none.
5 of 6 sections reviewed
Sign note
Signing unlocks only after every section is reviewed.
05 — HOW I'D MEASURE IT

What success would look like

North star: documentation time per visit, and the share of notes finished the same day instead of late at night. Adoption: weekly use by nurses, without a mandate.

Guardrail: the share of drafts nurses edit before signing. A high edit rate would be a good sign here — it means nurses are reading, not rubber-stamping. These are goals for a concept, not measured results.

06 — REFLECTION

What I'd carry forward

With AI, design the review, not just the output. The generation was the easy part. Making errors easy to spot was where design made the difference.

Plan for when the model is wrong. I'd build a lightweight "this is wrong" report into v1. It protects nurses, and it becomes the best source of training data.