Sifartek
A doctor in a white coat running an ultrasound scan
All workHealthcare · client project

ThyroidAI

Drafts the thyroid ultrasound report for an endocrinologist, and learns from every correction the doctor makes.

Client
A private endocrinology practice
Industry
Healthcare
Timeline
Since February 2026, in daily use
What we did
AI in operations, custom software

The problem

After every thyroid scan, the doctor writes the same structured report: the gland, each nodule with its features and EU-TIRADS class, a conclusion, and what to do next. The time goes into typing, not into the decision.

A general AI model can write that report in seconds. It can also invent a nodule that is not there. In medicine, that is not a small error.

How it works

  1. 01code

    Images and context

    Up to eight scan images and the clinical indication.

  2. 02AI

    The AI proposes findings

    Each lobe and each nodule, with sizes read from the doctor’s own calipers.

  3. 03person

    The doctor confirms

    Accepts or edits each finding before anything is written.

  4. 04code

    Code writes the report

    Classification, conclusion and plan are computed from the confirmed findings.

  5. 05code

    Corrections become rules

    Every edit can become a rule applied to every future report.

Inside the tool

Illustrations of how it works, drawn by us. No screenshots and no patient data.

Nodule 1, proposedillustration · no patient
  • LocationRight lobe, lower pole
  • Size14 × 9 mmfrom calipers
  • CompositionSolid
  • EchogenicityHypoechoic
  • MarginsRegular
  • Echogenic fociNone
EU-TIRADS 4computed by codeEditConfirm
The AI proposes, the doctor confirms. The classification is computed by code from the confirmed findings, so the table, the conclusion and the plan can never disagree.
Corrections564 corrections · 149 rules

The doctor’s correction

“Never write that the gland looks normal when the thyroid contains nodules.”

Permanent rule

Applied to every report from now on

Term replaced

Always “maladie de Basedow”, never “maladie de Graves”

A correction becomes a rule in one click. The doctor can switch any rule off at any time.
Reports to reviewheld by the verifier
  • A verifier rule was not metThe draft stops here until the doctor looks at it.
  • Image quality too lowThe measurements cannot be read reliably, so nothing is guessed.
When a check fails or an image is too poor to read, the report is held back instead of guessed.

Where the AI is, and where the code is.

The model does the reading. Code checks everything it produces before anyone relies on it.

AI does

  • Reads the scan images and proposes each finding.
  • Reads the measurements from the doctor’s calipers on the image.
  • Writes in the clinic’s own style and vocabulary.

{ code does }

  • Computes the EU-TIRADS class, the conclusion and the plan from confirmed findings only.
  • Holds a report back when a verifier rule fails or the image quality is too low.
  • Applies every permanent rule the doctor has set, on every report.
  • Keeps the doctor’s confirmation as a required step. It cannot be skipped.

So far

reports drafted in daily practice
214
corrections captured from the doctor
564
of them now permanent rules
149

ThyroidAI drafts documents for a specialist to check and sign. It does not make a diagnosis.

Next project

Clause

StartFree 30 minute call

Tell us what slows your team down.

Walk us through the process. We will tell you on the call what we would build, what it would cost, and whether it is worth it.

Or pick what you need

AI in our operationsConnecting our tools and dataAI inside our productCustom softwareNot sure yet
First call
Free, 30 minutes
Pilot running
In three weeks
Price
Fixed, agreed first

Reply within 1 business day

Registered in Estonia, EU