Bilal Shihab / Projects

Stealth Clinical AI

The on-device AI for a clinical startup I'm building, currently in stealth. It records a clinical conversation and analyzes it entirely on an iPhone:

FounderBuilt with a clinical partnerJun 2026 – present
Specmeasured on iPhone 17
Model
Qwen 3.5-4B, Q4_K_M, 3.0 GB
Runtime
llama.cpp on Metal GPU
Active RAM
< 500 MB (memory-mapped weights)
Full analysis
~144 s mean
Transcription
Apple SpeechAnalyzer, 3.1% WER
Model selection
240 labeled decisions
Second pass
65.1% → 84.1% agreement on synthetic conversations
Leaves the phone
Only reviewed scores, if the user opts in
The app's recording screen
Fig. 1The recording screen. Recording, transcription and analysis all happen on the phone.

Why I built it

Doctors I've talked to told me how much of their time goes into reviewing clinical conversations. I wanted to see if a phone could take on part of that without sending anything to the cloud.

The hard part

Most of my time went into choosing the model. In an early test, MedGemma 4B, Google's medical model, gave unearned credit 53% of the time. Apple's built-in Foundation Models require no download, but they matched my hand grading only 48.4% of the time, compared with 65.1% for Qwen 3.5-4B on the same conversations. On a larger set of 240 labeled decisions, Qwen 3.5-4B reached 85% against 79% for a 7B model, with a 3.0 GB download.

Qwen 3.5-4B 85.4% Qwen 2.5-7B 79.2% Qwen 3-4B (2507) 72.9% Qwen 2.5-3B 70.8% 0 50 100%
Fig. 2Agreement with 240 labeled decisions, every model run through the same scoring code. I shipped Qwen 3.5-4B. It never incorrectly denied credit (0 of 128); all of its errors were over-credits, which the second pass is designed to catch.
Writing: three speech engines, and why I shipped the least accurate oneWriting: Core ML vs Core AI for running an LLM on an iPhoneWriting: Apple's on-device model vs Qwen 2.5-7B