Skip to content

Projects Work Search

zenz-coreml

On-device multimodal Core ML experiment

Category: ML
Year: 2024 - 2025
Built with: Core ML

Timeline

2024 - 2025

General model conversion successful
Stateful model conversion successful
On-device porting experiments in progress

Tools & skills

Skills

Core ML / MLCompute
Swift
On-device inference pipeline configuration

Motivation

I wanted to build ML models optimized for Apple CPU, GPU, and NPU that could be used in mobile keyboards.
I wanted to implement AI-powered autocomplete inference for keyboards, given their limited memory resources and the fact that CoreML doesn't track memory.

What I learned

Learned how to convert PyTorch models using coremltools.
Studied how to convert to Stateful models using the KV-cache introduced in iOS 18.

Related projects

CASK

LLM · KV cache

CASK studies KV cache compression during long LLM reasoning. It keeps answer-critical states and merges repeated intermediate states. Published on arXiv on April 13, 2026 (arXiv:2604.10900).

Pylibkit

Swift · Python

A toolkit that combines SwiftPM templates, a Python CLI, and signing automation so Apple-platform projects can bundle Python runtimes and packages with less setup.

FloatingHUD

A reusable pure-SwiftUI HUD package with draggable, edge-snapping compact and expanded states for metrics and debugging overlays.