Product-Safety Diagnostics for a Medical Group
Netherlands
An AI system that reads twenty thousand product records, extracts their ingredients, and tells a clinician whether a given product is safe for a given patient.
- Client
- A Dutch medical-diagnostics group
- Industry
- Healthcare
- Period
- Jan 2025 — Sep 2025
Outcomes
20,000+
Products analysed
9 months
Engagement
- Client certification
- ISO 27001 și ISO 9001
The situation
A network of allergy and dermatology clinics needed to tell a patient whether a given consumer product was safe for them. The ingredient data existed across manufacturer sites, safety data sheets and product labelling, in no consistent format.
What we built
Acquisition
A scraper suite over retail sources, recording what was fetched and when. A clinical recommendation has to trace back to a source document, so provenance is stored with every record.
Extraction
Multi-modal extraction against declared Pydantic schemas. Safety data sheets arrive as scanned tables, spreadsheets with merged cells, and photographs, so output is a validated record rather than text that still needs reading.
Diagnosis
Retrieval over chemical and allergen context, then an assessment against a patient profile with its supporting evidence attached. Two generations: allergenic profiles first, chemical composition second.
Evaluation
A generic expected-versus-predicted comparison over structured output. This is the component that made the rest improvable — it turned "the model seems better" into a measurement.
What mattered
Retrieval was the bottleneck, not the language model. Getting the right chemical context in front of the model moved accuracy considerably further than changing model provider did.
Outcome
Over 20,000 products analysed, with per-patient assessment and alternative recommendation. Nine months, ending with the system in the client’s hands.