Experience

Magnosco builds computer-vision-assisted diagnostic tools for skin cancer detection, working closely with dermatologists and imaging scientists. My role is bringing their AI research into a production Apple application — not training the models, but turning a research-grade imaging device into a product a clinician can trust, from the Wi-Fi connection through image preprocessing, on-device inference, and secure transmission of results.

Challenges

• Integrating a dermatological imaging device over a device-hosted Wi-Fi network with no internet path, handling reconnection and firmware drift gracefully.
• Deploying and version-managing multiple CoreML models on-device with strict latency and thermal budgets.
• Building an OpenCV-based preprocessing pipeline that turns raw sensor output into consistent, model-ready imagery.
• Redesigning a memory-heavy computer vision workflow on iPad with actor-based concurrency, an intelligent caching strategy, and conditional model execution — reducing peak memory from roughly 4GB to 1.2GB and eliminating the crashes that came with it.
• Securing every request between the app and backend with mutual TLS, including on-device certificate lifecycle management.

Technologies

Impact

• Translated PyTorch model logic into efficient, production-ready Swift implementations used in a clinical setting.
• Cut peak memory usage on iPad from roughly 4GB to 1.2GB and eliminated device crashes through actor-based concurrency and intelligent caching.
• Reduced end-to-end capture-to-result time by streamlining the imaging and preprocessing pipeline.
• Established the app's networking and device-pairing layer as the reliability foundation the rest of the product builds on.