Magnosco — Diagnostic Imaging Platform

Problem

Magnosco's research team had validated a computer vision approach to assist dermatologists in early skin cancer detection, developed and trained in PyTorch. The gap was turning that research into a product a clinician could rely on in an exam room — a reliable device connection, consistent image capture, on-device inference within clinical time constraints, and a secure path for results.

Approach

I built the iOS layer that takes the device from research prototype to clinical tool: a resilient Wi-Fi pairing flow to the imaging hardware, an OpenCV preprocessing pipeline that normalizes captures regardless of lighting or device firmware version, and a CoreML inference pipeline running multiple models translated from the team's PyTorch implementations. Every result is transmitted over mutual TLS, with certificates provisioned and rotated on-device.

Architecture

The app is structured around a capture-to-result pipeline: a device-communication layer over the local Wi-Fi network, an image-processing module built on OpenCV and Core Image, a CoreML inference layer managing multiple versioned models, and a networking layer enforcing mTLS for every request. Each stage is isolated behind a protocol so models and preprocessing steps can evolve independently as the science improves.

Technical Decisions

Challenges

• Translating PyTorch model logic into efficient, numerically consistent Swift/CoreML implementations.
• Handling device pairing and reconnection gracefully on a Wi-Fi network with no internet path.
• Meeting the accuracy and consistency bar required for a clinical support tool, not just a demo.
• Redesigning a memory-heavy computer vision workflow on iPad that was causing crashes under sustained use.

Results

• Cut peak memory usage on iPad from roughly 4GB to 1.2GB and eliminated the device crashes that came with it.
• Shipped a production pipeline translating the research team's PyTorch models into Swift/CoreML without a loss of clinical accuracy.
• Built the device-pairing and mTLS networking layer clinicians rely on during live examinations.

Lessons learned

• The hardest part of shipping AI in a real product is rarely the model — it's the pipeline of unglamorous engineering around it.
• Working directly with scientists sharpened how I reason about model behavior, uncertainty, and where engineering judgment should and shouldn't override research decisions.
• Medical-grade reliability requirements make you a better engineer everywhere else, too.