Back home

The Last Mile of the Fax - Split, Redact, and Annotate

After Fax Transform shipped, one truth emerged: the magic wasn't in extraction — it was in what clinicians did next. I owned the second act end-to-end: the review surface where multi-patient faxes get split into correct records, sensitive data gets redacted before it touches the chart, and clinical signal gets annotated onto the artifact so nothing is ever re-read twice.

1

Designer, owned end-to-end

~1

Hour/day saved per reviewer

47

shadow hours at live review desks
  • Split
  • ·
  • Redact
  • ·
  • Annotate
  • ·
  • HIPAA-grade
  • ·
  • Enterprise-Ready
  • ·
01 · Overview
The demo worked.
The workflow didn't.
That was my brief.
Role

Sole Product Designer · end-to-end

Scope

Research · IA · Interaction · Visual · QA

Platform

Web · tablet stylus · keyboard-first

Partners

3 engineers · 1 PM · clinical advisor

Outcome

First enterprise contracts · 2 health systems

02 · problem
A single fax isn't a document. It's five patients' records printed as one.
Extraction got us to 97% field accuracy. The remaining mile — figuring out where one patient ends and the next begins, which boxes hide a Social Security number, which highlighted line is the reason the fax was sent at all — was still being done by hand.

The early Fax Transform review surface was a PDF viewer with a sidebar. It shipped because it worked for one patient per fax. In practice, a significant share of inbound faxes bundled two or more patients, and nearly all contained at least one field that had to be redacted before the document could be attached to a chart. The "last mile" was actually the whole marathon.

01
2.3 min average review time per page — up from 40s in the demo
02
Zero tooling for redaction — staff printed, blacked out with Sharpie, re-scanned
03
Zero tooling for annotation — staff have to use internal communication tool
“Your app is beautiful. I still have a Sharpie in my other hand because of page four.”

— Priya M., Medical Assistant · Horizon Family Care · review-session day 3

03 · RESEARCH
I watched 47 hours of review before I opened Figma.
A
Live review shadowing

47 hours across clinics and regional health systems. Every click, every undo, every "wait, go back" logged with timestamps. I built the interaction map before I built the screen.

A
Live review shadowing

47 hours across clinics and regional health systems. Every click, every undo, every "wait, go back" logged with timestamps. I built the interaction map before I built the screen.

B
Task decomposition

Broke "review a fax" into discrete micro-tasks. Mapped each to tool and frequency. The split and redact steps dominated — everything else was overhead.

B
Task decomposition

Broke "review a fax" into discrete micro-tasks. Mapped each to tool and frequency. The split and redact steps dominated — everything else was overhead.

C
Redaction liability audit

Partnered with the clinical advisor and an outside HIPAA attorney to understand what had to be redacted, and what happened when it wasn't. The answer shaped the entire redaction flow.

C
Redaction liability audit

Partnered with the clinical advisor and an outside HIPAA attorney to understand what had to be redacted, and what happened when it wasn't. The answer shaped the entire redaction flow.

D
Prototype Testing

Tested multiple approaches for split — page-cutting interactions, checkbox selection, range input. Watched where users hesitated. That hesitation became the design constraint.

D
Prototype Testing

Tested multiple approaches for split — page-cutting interactions, checkbox selection, range input. Watched where users hesitated. That hesitation became the design constraint.

04 · insights
Three craft decisions reframed what we were building.
01
Split is an assignment tool, not a cutting tool.

Early thinking framed split as dividing a document — placing cuts between pages. What reviewers actually needed was to assign pages to patients. "Pages 1–3 belong to Eric Nam, pages 2–3 belong to Mark Whitaker." The mental model shifted from scissors to a roster. The design followed.

02
Redaction should be manual, never automated.

A mis-redaction isn't a bug — it's a compliance incident. We don't surface AI suggestions or auto-redact candidates. The reviewer adds each redaction themselves, specifies the page, and optionally describes what was redacted. The human is the last signature, always.

03
The confirmation step is part of the design, not a formality.

Before any split is processed, the reviewer sees a summary: every fax, every patient, every page range, every PHI warning — in one modal. This wasn't a safeguard bolted on at the end. It's where the work becomes legible. The moment a reviewer can scan the whole thing and say "yes, that's right" is the moment the tool earns their trust.

05 · solution
One queue.
One workspace.
No jumping between tools.

The queue shows every incoming fax with patient match status, document type, confidence score, and who last touched it. A reviewer opens a fax and sees it whole — the document on the left, the work on the right. They assign pages to patients, flag any PHI, and confirm before anything moves to the chart. No context switches. No separate redaction tool. No printing.

  • Split view
  • redact view
  • Split view
  • Split view
  • redact view
  • Split view
Craft principles
01
The queue tells you where to start.

Multi-patient faxes are flagged before a reviewer opens them. Priority and confidence score surface at a glance. Reviewers don't triage from scratch — the system does the rough sort, the human makes the call.

01
The queue tells you where to start.

Multi-patient faxes are flagged before a reviewer opens them. Priority and confidence score surface at a glance. Reviewers don't triage from scratch — the system does the rough sort, the human makes the call.

02
Page assignment, not page cutting.

The split panel shows which pages belong to which patient with overlapping ranges allowed — because the same page can contain two patients' information. You assign, not divide.

02
Page assignment, not page cutting.

The split panel shows which pages belong to which patient with overlapping ranges allowed — because the same page can contain two patients' information. You assign, not divide.

03
Redaction is per-fax, not per-document.

After splitting, redaction happens on each resulting fax individually — with the source document visible on the left. Reviewers see exactly what the next person will see before they confirm.

03
Redaction is per-fax, not per-document.

After splitting, redaction happens on each resulting fax individually — with the source document visible on the left. Reviewers see exactly what the next person will see before they confirm.

04
Confirmation before consequence.

The confirm modal shows every fax, patient, page range, and PHI warning before anything is processed. It's the last moment to catch an error. Making it readable — not just a checkbox — was a deliberate choice.

04
Confirmation before consequence.

The confirm modal shows every fax, patient, page range, and PHI warning before anything is processed. It's the last moment to catch an error. Making it readable — not just a checkbox — was a deliberate choice.

06 · the flows
Three tools. Two flows. One direction. Nothing processed until it's right.

Each flow was prototyped and tested with clinicians against one benchmark: could they process a multi-patient fax faster and more accurately than the paper-and-Sharpie method it replaced?

Flow 01 · Split
01
Open the fax

The reviewer opens a fax from the queue. Multi-patient faxes are already flagged. The document renders on the left; the right panel shows the current page and two options: add it to a split fax, or archive it.

02
Assign pages to patients

For each split fax, the reviewer selects the patient from a dropdown and sets the page range. Overlapping pages are allowed. The system flags "Multiple patient detected" where it finds shared content — the reviewer confirms or corrects.

03
Confirm the split

A modal summarises every resulting fax: patient name, document type, page range, and any PHI warnings. The reviewer can go back or confirm and move to redaction.

Flow 02 · Redact & Annotate
01
Review each fax

After splitting, each resulting fax opens for review in sequence. The source document stays on the left. The right panel shows the patient, document type, a redaction log, and an annotation area — both available at the same time, in the same pass.

02
Redact PHI

The reviewer adds redactions manually: page number and an optional description of what was removed. Each redaction is logged. Faxes with no PHI to redact can be skipped.

03
Annotate as needed

In the same panel, the reviewer can leave annotations — notes for whoever picks up the chart next. Redaction and annotation happen together, without switching screens or tools. One pass covers both.

04
Process the split

On the final fax, "Next fax" becomes "Process the split." Nothing moves to the chart until the reviewer makes that call. Everything — splits, redactions, annotations — is confirmed together.

07 · impact
The flow that made Fax Transform enterprise-ready.

The original product shipped as a clinic tool. Split, Redact, and Annotate was the work that let us walk into health-system procurement, survive legal review, and sign our first two enterprise contracts.

Craft principles

0

-

Reported PHI leaks across 410,000 reviewed faxes in the first 90 days of production.

1

hr

Enterprise contracts signed in the quarter after launch. Both cited the audit trail by name.

2

×

Enterprise contracts signed in the quarter after launch. Both cited the audit trail by name.

75

%

Reduction in time-to-review on multi-patient faxes. From around 2 min per page to 30 seconds.

08 · reflection
What the craft taught me.
The confirmation step is a design surface, not a checkbox.

The moment I started treating the confirm modal as a trust-building moment — not a legal formality — the whole flow clicked. Reviewers needed to see their work summarised before they committed to it. That's a product insight, not just a UX detail.

Audit is a design material.

The redaction log — page number, description, reviewer — wasn't a backend artifact. It was the thing that moved procurement from "maybe" to "yes." Treating it as a first-class product surface, visible to the reviewer before they submit, changed how legal thought about the tool.

One designer can own a flow end-to-end, if the flow is small enough.

Split, Redact, Annotate — three verbs. Owning all of it, from research to QA, meant the micro-decisions stayed coherent. The next flow should still be one designer, but I'd carve it smaller next time.