Stealth Clinical AI

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

on-device transcription, at a 3.1% word error rate
Qwen 3.5-4B for the analysis, with a second pass that re-checks its work
about two and a half minutes per analysis, in under 500 MB of memory
nothing leaves the phone

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: Verifier, 65.1% → 84.1% agreement
Leaves the phone: Nothing without review

The app's recording screen — **Fig. 1**The recording screen. The audio and the analysis stay 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

Choosing the model. MedGemma 4B, Google's medical model, gave credit that hadn't been earned 53% of the time. Apple's built-in Foundation Models needed no download but agreed with my scores only 48.4% of the time. Qwen 3.5-4B reached 85% on 240 labeled decisions, beating a 7B at 79%, in a 3.0 GB download.

Fig. 2Agreement with 240 labeled decisions, every model run through the same scoring code. I shipped Qwen 3.5-4B. It never wrongly denied credit (0 of 128); all of its errors were over-credits, which is what the second pass goes after.

Writing: three speech engines, and why I shipped the least accurate one Writing: Core ML vs Core AI for running an LLM on an iPhone Writing: Apple's on-device model vs Qwen 2.5-7B

Next projectLocalabs →