Localabs

An iPhone app that turns a photo of a lab report into a plain-language explanation, with a medical language model running entirely on the phone:

Scans multi-page reports and reads them with Apple's on-device OCR
MedGemma 4B writes a summary, a glossary, diet notes, questions for your next visit and a list of the medications on the report
Turns those medications into a daily reminder list
Lets you circle any part of the scan and ask about it
No account and no cloud

Solo projectApr 2026 – presentFeature-complete, not on the App Store

Specmeasured on iPhone 17

Model: MedGemma 4B, 2.5 GB quantized
Runtime: llama.cpp on Metal GPU
Photo to readout: ~12 s
OCR: Apple Vision, table reconstruction
Output: Summary, glossary, diet, questions, medications
Context: 30 days of HealthKit data
Cloud baseline: Command A+: 7.1 s median, ~$0.02/report
Requires: iOS 26, iPhone 15 Pro or later

Three Localabs screens: a report summary, the medication list, and questions for the doctor — **Fig. 1**A report summary over the original scan, the medication list it builds, and the questions it suggests for the next visit. The report is a sample.

Why I built it

Lab results are hard to interpret, especially when no one at home has medical training. A general-purpose chatbot can hallucinate, so it shouldn't be relied on for this, and it won't remind you about the medications on your report or suggest questions for your follow-up visit.

The hard part

The main challenge was keeping everything on the phone, including iPhones with less memory. With the 2.5 GB model loaded, processing a full 12-megapixel photo pushed 6 GB iPhones past their memory limit, and iOS shut the app down mid-scan. I fixed this by downscaling each photo to 2048 px before OCR. The app also checks for at least 7 GB of RAM at launch, because the App Store can't restrict downloads by device memory.

localabs.app Code on GitHub Cloud benchmark

Next projectScribend →