Healthcare · optometry · HIPAA · Wichita Family Vision · Bedrock live 2026-09-16

Invoice filing for a clinic's bookkeeper: HIPAA-compliant on Claude via AWS Bedrock, at 0.9 cents an invoice.

The bookkeeper keyed every vendor bill by hand and filed PDFs into 184 folders. Now a vision model reads each one, proposes the folder, and waits for her click, on a HIPAA-compliant path.

0.9¢ per invoice, 3.4 seconds a file, measured on Claude Sonnet 4.6 in the clinic's account 26,904 input and 2,003 output tokens for 12 files at Bedrock on-demand pricing: $0.11 a run.

12 of 12 synthetic invoices read correctly on 2 models, rotated scans included

The review table: 12 files, proposed names and folders, vendor read, date, amount, and a note per row — The review page on 12 synthetic invoices. Amounts are invented; nothing moves until Go.

At a glance

Client: Wichita Family Vision, an independent optometry practice in Wichita, Kansas.
Who uses it: Built for the clinic's bookkeeper. Tested end to end in the clinic's own AWS account on 2026-09-16; her desk install is the next step.
Shipped: Bedrock live 2026-09-16. About 4 weeks, including the compliance re-architecture.
Stack: Claude on AWS Bedrock
Python
boto3
PyMuPDF
QuickBooks Online bill capture
Windows
Compliance: HIPAA-compliant by design. Business associate agreements in force with the clinic and with AWS; the model runs in the clinic's own AWS account with retention off; patient-named documents never reach QuickBooks or a chat window.
Cost to run: 0.9 cents an invoice on Sonnet 4.6, about 0.3 on Haiku 4.5. At any plausible clinic volume that is $2 to $13 a month of AWS.
My role: I built it: the runner, the Bedrock path and the BAA setup in the clinic's account, the compliance finding that moved the architecture, and the synthetic test set. A second engineer ran the QuickBooks trial tests.

What I built

The build

The accounting entry without AI. Approved PDFs flow into QuickBooks bill capture, which fills the bill from the vendor's default category. No model touches the ledger.
A vision-model filer for the messy input: a program on her PC reads each PDF (rotated scans included), matches it to the vendor folder, proposes the new file name, and shows a review table. Nothing moves until she clicks Go.
Every move logged and undoable. Unknown vendors are flagged for her instead of guessed. Patient-named documents are filed but never queued for QuickBooks, and their names never enter the log.
The model call runs on Claude through AWS Bedrock inside the clinic's own AWS account, under the standard AWS BAA, with data retention set to none.

Diagram of the filing flow from PDF to QuickBooks — Read, match, propose, review, move, queue.

What changed

Before and after

Before

By hand

Reading a billEyes, then keyboard

FilingHand-filed into 184 folders

The accounting entryTyped into QuickBooks

Patient-named documentsSame pile as everything else

Cost per invoiceBookkeeper minutes

After

The system

Reading a billVision model, 3.4 s a file

FilingProposed, reviewed, one click

The accounting entryQuickBooks capture, no AI

Patient-named documentsFlagged, filed, never sent on

Cost per invoice0.9 cents of model time

Where the numbers come from. Measured on 2026-09-16 in the clinic's AWS account: 12 synthetic invoices, 26,904 input and 2,003 output tokens on Claude Sonnet 4.6 at $3 and $15 per million, $0.11 a run, 0.9 cents an invoice, 3.4 seconds a file. The same 12 files read identically on Haiku 4.5 at about 0.3 cents. The hours saved will be measured on her real volume after the install.

For engineers

How it's built

Architecture: A local Python runner: PDF pages to PNG at 150 DPI (3 pages max), one prompt with two backends (Bedrock for production, OpenRouter for synthetic tests only), a deterministic vendor matcher (exceptions, normalized exact, unique containment), a plan step that names files and resolves collisions, a review page on 127.0.0.1, then move, log, and queue for QuickBooks.
Where it runs: On the bookkeeper's PC against the clinic's network drive. Only the reading step leaves the building, to Bedrock in the clinic's own AWS account (us-east-1), under the AWS BAA. Stephens AI operates it through an access key scoped to one inference profile.
Tests and evals: 12 synthetic invoices mirroring her real nine plus the untested cases: 5 rotated image-only scans, a bare-date receipt, a paid invoice, a statement with a period end, an unknown vendor, a same-name collision, 2 patient-data documents. 12 of 12 on Sonnet and on Haiku. Full run, undo, and the browser Go path all proven.
Cost and latency: 0.9 cents an invoice on Sonnet 4.6, about 0.3 on Haiku 4.5. At any plausible clinic volume that is $2 to $13 a month of AWS.
Guardrails: HIPAA: a business associate agreement between Stephens AI and the clinic, the AWS BAA on the clinic's own account, retention set to none in Bedrock, logging off, the Anthropic use-case form filed. No patient name is ever written to the log or the proposal. Nothing is deleted or overwritten, ever; the tool never creates a folder; a stop-on-first-failure move with a full undo.
The tradeoff we chose: Why this way and not the other
We moved off Claude Team and Cowork mid-build. They are the easier surface, and the first version ran there, but neither carries a BAA, so the runtime could never read a patient-named file. The Bedrock path costs a compliance program and a Windows installer; it is the only path where the clinic's data stays inside its own boundary. Also chosen: QuickBooks' own capture for the accounting entry, because a model should not touch a ledger when a deterministic feature already does it.