SB/serdar

AI & Intelligent Products

I build AI-powered Apple products — I don't train the models.

My work sits between research and production: taking a validated model or a well-scoped LLM feature and turning it into something reliable, fast, and honest about what it does with a user's data.

At Magnosco, I work alongside scientists building computer vision solutions for skin cancer diagnostics — integrating multiple CoreML models into a production iOS application, building the OpenCV pipelines that feed them, and translating PyTorch model logic into efficient Swift. I am not an AI researcher, and I don't train foundation models. What I specialize in is the harder-than-it-looks work of bringing AI technologies into a production-quality Apple product — reliably, on-device, and with privacy treated as a starting constraint rather than an afterthought.

Where I work

From research prototype to production pipeline

CoreML

Deploying, versioning, and optimizing trained models on-device — managing compute unit selection, quantization, and the lifecycle of multiple production models at once.

OpenCV

Building preprocessing and feature-extraction pipelines that turn raw camera or sensor input into consistent, model-ready data in real time.

Medical Imaging

Working directly with imaging scientists to turn validated research into a pipeline that holds up to the accuracy and consistency bar a clinical tool requires.

Image Processing

Color space normalization, orientation correction, and region-of-interest extraction — the unglamorous steps that determine whether a model's output is trustworthy.

Feature Extraction

Structuring the path from raw pixels to the specific signal a model was trained on, tested independently from the model itself.

Edge AI

Designing inference to run entirely on-device — for latency, for reliability without a network connection, and because it keeps sensitive data off a server entirely.

On-device Inference

Profiling models for real-world thermal and battery budgets, not just accuracy — the difference between a demo and a feature people use all day.

PyTorch Collaboration

Translating a research team's PyTorch model logic into efficient, numerically faithful Swift and CoreML implementations, with tooling to catch drift early.

LLM Integration

Scoping language-model features narrowly around specific tasks, favoring on-device foundation models where possible, and being explicit with users about what data a prompt contains.

Prompt Architecture

Treating prompt templates as product surface, not implementation detail — versioned, reviewed, and designed with constrained output formats and sensible fallbacks for when a model response doesn't parse.

Claude API Integration

Wiring Claude into narrowly-scoped product features via the Messages API — structured prompts, minimal payloads, and no server-side retention of conversation history by default.

Privacy-first AI

Treating every AI feature's data flow with the same scrutiny as any other privacy-sensitive code path — minimal payloads, no server-side retention by default, honest UI copy.

Case study

See it in a real product

The full story of Magnosco's diagnostic imaging pipeline — device integration, CoreML, and mTLS networking.

Read the case study