Comparing store names with == fails almost every receipt. Here are the three code metrics and the model judge that turn a photo extraction into numbers, plus the results on 21 real receipts and the error patterns hiding behind them.
Multimodal prompting collapses OCR and understanding into one call. Here is the feature: a @Generable Receipt, one carefully shaped session with an image Attachment, and the photo dataset that has to be transcribed by hand before anything can be measured.
Two refinements to the NeighborhoodCoherence judge from part two: pass the metric name straight to ModelJudgeEvaluator to drop the ScoreDimension boilerplate, and shape what the judge actually reads with evaluationTarget and reference.
Code-based evaluators can only measure what code can measure. This companion article goes deep on ModelJudgeEvaluator: choosing a scoring scale, writing scoring levels that actually work, and wiring a model-as-judge into the TripPlanner evaluation from part one.
How to test generative, non-deterministic features with Apple’s Evaluations framework, the five steps of an evaluation, datasets and the Loader protocol, synthetic sample generation, evaluators and metrics, and the compiler gotchas hit along the way.
A walkthrough of Anthropic’s new Claude Agents (agent view) inside Claude Code on iOS: how to run a Clean Architecture refactor agent and a unit-testing agent in parallel on isolated Git worktrees, with the full lifecycle and synchronisation pitfalls.
In Part 1 of this series, we built a minimal LLM inference pipeline on Apple Silicon using MLX Swift. In Part 2, we quantized a model from scratch and saw how 4-bit precision makes billion-parameter models tractable on a phone. This article introduces MLX Embedders, specifically the MLXEmbedders Swift library, and takes a different angle. Instead of generating text, we are going to encode meaning. A user types: “best hikes near a volcano.” A keyword search returns nothing useful. An embedding-based system returns exactly what they need, because it understands what the words mean, not just what they spell. That capability is what embeddings unlock, and it is the foundation of every serious AI feature in production today: semantic search, recommendation systems, RAG pipelines, and clustering. ...
Abstract The proliferation of large-scale neural language models has, until recently, been contingent upon access to remote computational infrastructure. The architectural characteristics of Apple Silicon, most notably its unified memory subsystem, present a substantive departure from this dependency. This article examines MLX Swift, a native Swift binding to Apple’s MLX machine learning framework, as a mechanism for deploying quantized Large Language Models (LLMs) directly on consumer Apple hardware. We characterize the layered architecture of the MLX ecosystem, contrast its design philosophy with that of Apple’s Foundation Models API, and present a reference implementation demonstrating the complete inference lifecycle: model acquisition, session initialization, and autoregressive text generation. The discussion is grounded in the computational properties of unified memory and their implications for on-device inference efficiency. ...
The landscape of iOS development has dramatically shifted in 2025. With Apple’s introduction of Xcode 26 at WWDC 2025, which integrates ChatGPT and supports multiple AI models through API keys, and Anthropic’s release of Claude Code as a powerful command-line tool for agentic coding, developers now have unprecedented AI-powered development capabilities. ...
Swift’s concurrency model introduces actors to provide safe, isolated access to mutable state. However, constantly using await when interacting with actors can create performance bottlenecks and reduce code readability. The isolated keyword offers an elegant solution by allowing synchronous access to actor state in specific contexts. ...