Doctor profiles took an hour each by hand. Now an agent writes 25 an hour, with sources

We went through the patient’s path with the client in short rounds, one screen at a time. For each screen we asked what the patient is deciding there. We kept changing it until the page was consistent, quick to read and easy to move through. The audit showed what was broken. These sessions decided what replaced it.

Colors, type sizes and spacing are defined once, in one file. Every component takes its values from there, so a change to the brand color reaches every page. The redesign was delivered as a working site of 25 pages that needs no framework to run. It was not a design file that someone still had to build.

Every piece of text in the Figma design was compared automatically with the text on the built site. Sizes and positions were measured in a browser, not judged by eye. Before fixing a difference, we checked whether it was a later change the client had asked for, so the check could not undo it. Errors were fixed on whichever side had them, including in the design file.

Before the agent ran, we wrote the rules every profile must follow. Examples: titles and names are written in Latin letters, a doctor’s main specialty comes only from the current hospital’s page, hospitals are always named in English, and a year is never guessed. The client can read each rule and dispute it.

The agent takes a doctor and searches public sources: the hospital’s page, PubMed, ORCID, Google Scholar, congress pages. It confirms each source is the same doctor by comparing name, photo, hospital, city, specialty and education, and gives each match a confidence level. It writes each field straight into the form the rules set. Low-confidence findings go into notes for the editor, not into the profile. Two doctors with the same name are never merged automatically.

After research, a separate pass challenges each claim before it is accepted. In one profile, a publication under a different surname was left out because no source confirmed the name change. In the same profile, an experience figure worked back from a conference date was left out. The reason for each is written into the editor’s notes.

Each profile is laid out as a Word document for the content manager to review. While it builds the document, the software checks eight of the sixteen rules, such as no pronouns in the biography and no “present” in a CV line. Each violation prints a warning with the doctor’s name, and every run ends with a count of warnings.

Nobody retypes a finished profile. A plugin we wrote for the marketplace imports each record with its fields, categories and page addresses already filled in. Photos stay empty on purpose, and the content manager adds them.