Watch the project video (2m 53s)
'Umbra' is a proof-of-concept privacy-first tool that redacts Personally Identifiable Information (PII) from text information on a local device using Gemini Nano AI. Redacted texts can then be sent to Large Language Models and AI cloud services for generative tasks. When a response is received, Umbra locally restores the redacted data back into the text information.
The tool has many potential. For demonstration, we chose the use case of a fictitious ‘Bravo Children’s Hospital’ (video).
Anonymised texts about the patient’s medical condition is used by Gemini to:
for the doctor: write a letter to the patient’s insurance company justifying the necessary treatments. The returned letter is repopulated with PII. The doctor checks and sends it off to the insurer, making their workflow more efficient.
for the child patient: write a story based on their medical condition, using their favorite toy as the central character. The story texts returned from Gemini can then be used by an image generation AI to create images. The images and story title are laid out on 1 page using a third party API. A nurse receives it as a PDF file, prints it using any desktop printer, and folds the printed sheet into a picture book following the instructions provided. The child leaves the hospital visit at the end with a personalized story book to help them understand their medical situation.
I worked with a senior backend developer on this project.
Technology:
Google Gemini Nano
Google Gemini Pro
Bulma CSS framework
Also:
Adobe Firefly
Canva