While redesigning Sonovit’s website, I found a second problem inside the clinic. After each hearing evaluation, results moved through phone photos, handwritten notes and manually formatted Word files. Identical studies could leave the clinic as visibly different reports.
I designed, implemented and deployed an internal tool that captures structured study data, reconstructs the charts and generates a standardized PDF for professional review. The final version launched across both branches in early July 2026.
Inputs arrived in incompatible formats — phone photos of the audiometer, scanned printouts, clinical notes as long paragraphs — and each professional rebuilt the report in their own way.
What was missing wasn’t a better template. The study had no structured form: thresholds existed only inside a photograph, so the values could not be validated, recalculated or compared.
One rule shaped the architecture: nothing reaches the patient document without first existing as structured, validatable data. That split the workflow into capture, extraction, interpretation and generation.
Clinical wording comes from deterministic rules, not a language model. AI may assist with reading source files and organizing notes, but never writes the conclusion — the professional approves before export. The trade-off: less generative flexibility, more consistency and clinical control.
I started by interviewing the physicians and hearing-care professionals who perform, interpret and review the studies. Those conversations defined which clinical information to capture, which tests to support, and the parameters and formulas behind each one.
I then piloted the platform for two weeks with four users across both branches, during real consultations — 15 reports before the final release. Their feedback simplified data capture, cut fields the report didn’t need, added autosave, and clarified how accounts and access permissions are managed.
I owned the workflow, information architecture, interface, implementation, testing and deployment. Clinical requirements — studies, parameters, formulas and report content — came from interviews with physicians and hearing-care professionals.
The tool was designed for clinical staff who shouldn’t need software training: one primary action per screen, clear progress, plain-language feedback and autosave throughout. The interface follows the consultation rather than asking the clinic to adapt to it.
1
2
3
4
All portfolio screens use synthetic demo data. No patient information is shown.
The difficult decisions were not about automating more. They were about deciding what could be standardized, what still required professional judgment and how the system should behave when information was uncertain.
A study can be complete for its scope even when some tests weren’t required. Empty sections are dropped from the patient document; missing-data warnings stay inside the tool, where they guide the professional.
Result — A complete report, not a visible list of omissions.
Extracted values keep their confidence level and source. Low-confidence points are highlighted and block validation until reviewed. The system never interpolates missing values to complete a curve.
Result — No unmeasured number reaches the report.
Audiograms are reconstructed as vectors from stored thresholds using standard symbols. The editor and PDF use the same geometry source.
Result — The preview and exported report always match.
Export remains in draft until a professional confirms the data, charts and interpretation. Editing a validated study automatically returns it to review.
Result — Automation supports professional judgment instead of replacing it.
The interface was only half of the product. The patient receives the PDF, so it needed the same hierarchy, consistency and clinical legibility as the application.
It matches Sonovit’s printed identity, uses vector charts and selectable text, and includes only sections holding real study data. A tonal-audiometry-only evaluation produces a complete two-page report, not one padded with “not evaluated” pages.
Fictional example patient. Open the full PDF →
The final version launched in early July 2026 and is now used several times per week across both branches, often more than once a day. Nobody rebuilds documents from scratch in Word — structured data, reconstructed charts and report rules live in the same workflow.
The report no longer needs to be manually reconstructed in Word.
Both branches now produce documents using the same structure, visual system and validation rules.
I designed, implemented, tested and deployed the product end to end with AI-assisted coding tools, against three constraints from the clinic.
Consultations can’t stop when the internet fails.
Files stay local to each branch.
Silent errors must never reach an exported report.
Replaceable, able to stay on site. With no model configured the full workflow still runs in manual mode.
Each branch runs the same stack on its own machine, available across the internal network — data, charts and PDF generation stay local, so a poor connection never stops a consultation. Trade-off: records are not synchronized between locations automatically.
The clinical requirements did not come from the model. Studies, parameters, formulas and report content came from the physicians and hearing-care professionals who perform and interpret them — and deterministic rules plus professional review decide what reaches the final report.
Still Sonovit-specific — some values need manual capture, and the rules, report structure and clinical configuration were built for this clinic rather than as a configurable product.
The report is what the patient keeps and what represents the clinic, so its hierarchy, charts and pagination needed the same design attention as the interface.
Shipping it also changed my judgment as a designer. Watching real consultations taught me more about required fields, error copy, permissions and autosave than reviewing the flow on my own screen ever could.
Measure completion time, corrections and adoption more systematically before adapting the product for workflows beyond Sonovit.