Case study · OrthoFlow

Patient presentation automation

The operating question

How do you turn uneven clinical inputs into a consistent patient presentation without hiding missing information or removing operator judgment?

SettingOrthodontic clinic workflow
InputsIntake PDF, photos, clinic template
OutputDeterministic PPTX and audit
ArchitectureLocal first and stateless
EvidenceImplementation record, no outcome claim

Clinical presentations are not a mail merge.

Intake forms omit fields. Photo sets arrive incomplete. Clinic templates change. A system that quietly fills the wrong slide is worse than one that stops and shows the gap.

Q1

Can the clinic use its active template?

The system introspects the template rather than depending on a manually maintained slide map.

Q2

What happens when an input is missing?

Required gaps remain visible. The system does not invent clinical content to make a deck appear complete.

Q3

Where does human judgment remain?

The operator confirms patient identity, reviews warnings, and decides whether the deck is ready for use.

Before · assembly work

Each case requires staff to gather, place, and check the same classes of material.

  1. Read the intake PDF
  2. Sort clinical photos
  3. Find the right template slides
  4. Check every gap by hand
Designed path · supervised assembly

The repeatable work becomes a fixed pipeline while judgment stays visible.

  1. Load the case inputs
  2. Resolve fields and photo views
  3. Confirm identity and warnings
  4. Deliver a deck with an audit

One local run. One reviewable output.

The browser supervises the run but never edits slides. Each case moves through a deterministic local pipeline. Working files are removed after delivery or cleanup.

Clinic input · PHI stays on the workstation
Case output · Deck, warnings, and audit
Source set
Intake

PDF, clinical photos, and the active template enter one run.

Template model
Introspect

The system reads the template structure and derives its slots.

Case resolution
Resolve

Extracted values and photo views map to explicit slots.

Human gate
Confirm

The operator confirms identity and reviews visible warnings.

Reviewable artifact
Deliver

A deterministic PPTX, error report, and audit are produced.

What remains local: patient inputs, template, extracted fields, images, and working files.

What never happens: cloud processing, silent gap filling, or browser-based slide editing.

The gap state is part of the product.

The system makes uncertainty legible rather than smoothing it over.

Patient identity is confirmed

The operator checks case identity before accepting the presentation.

Missing inputs remain visible

Required gaps surface as warnings rather than fabricated values.

The output is reproducible

The same resolved case state produces the same presentation contract.

Working data is ephemeral

Per-run files live in temporary storage and are wiped after delivery or cleanup.

OrthoFlow supervisor interface in an unloaded local state showing input drop zones and pipeline stages
Actual built interface · Focused crop

The local supervisor before a case is loaded. This focused crop contains no patient data and shows only the input boundary and pipeline controls. It does not establish clinic adoption, update performance, deployment status, or a business outcome.

What this record establishes

A local-first system was built with deterministic assembly, visible gap handling, operator confirmation, ephemeral run data, and a reviewable output contract. This record makes no claim about clinic adoption, labor saved, error reduction, or business results.

Design the failure state before the happy path.

For document workflows carrying clinical or operational risk, speed matters only after the system can show what it knows, what is missing, and where a person must decide.

Next case: AI enabling a semi-private airline