Local question
The question stays on the device
Product / Case Study
On-Device Health Evidence Assistant
A health AI that answers by quoting real medical research — and never sends your data anywhere.
STATUSArchitecture complete · Build in progress · Prototype on request
Cloud assistants can answer questions but send personal health data to a server. Privacy-first apps store locally but cannot meaningfully answer. This system combines private storage with useful, cited answers.
01
What we built
Runs entirely on the user’s device — no server, login or account
Retrieves and quotes medical literature with a tappable source
Returns offsets into real text rather than free prose, so it cannot output content absent from the source
Supports real reminder schedules without giving dosing advice
02
Why it is hard
Seven checks must pass; otherwise the system honestly reports insufficient evidence
A dedicated decoy test checks whether the AI uses evidence rather than memory
Evidence lookup and personalisation remain separated by design
03
How it works
Retrieve
Find relevant medical literature
Rank evidence
Select the strongest source passages
Extract offsets
Return exact passages, not free prose
7 checks
Run all seven evidence checks
Cited answer
Show the quote with its citation
04
Where it applies
Licensable evidence engine for health apps, insurers and patient support
Adaptable to legal, financial and internal knowledge systems
For any domain requiring quote, don’t guess
On-device AILocal encryptionEvidence retrievalOffset extractionVerification chainPrivacy by design
Have an operation that should work smarter?
We turn the right operational problem into a controlled AI product.