Nine public projects at github.com/nicedreamzapp, ordered by relevance to on-device ML and Apple Silicon work. Each links to a longer case study with file references, bugs from the commit history, and the numbers the repos actually contain.
Much of the 2026 work was written with Claude Code as a pair programmer; commit trailers show which commits. Upstream models and frameworks are credited in each case study.
A free iPhone and Android camera app for blind and low-vision users that detects 601 object classes with YOLOv8 on Core ML and describes scenes with a fine-tuned Qwen3.5-0.8B through MLX Swift, all on the phone. The work is in the parts that make it hold up: thermal tiers, a Neural Engine workaround, jetsam fixes, stride-correct tensor decoding, and an OCR cross-check on the model's dollar amounts.
A pure-MLX runtime for the vision and audio towers of NVIDIA's Nemotron-3-Nano-Omni, which previously ran on a Mac as text only. Each component is checked against NVIDIA's PyTorch reference with 23 parity tests (vision cosine 0.99996, audio 0.99999), and the backbone reuses mlx-lm without a fork.
A stdlib Python server that runs open-weights models through MLX and answers Anthropic's Claude Code CLI in its own API format, so Claude Code runs on a Mac with no cloud model. Fixes to KV cache trimming, prefill memory (34.2 GB to 20.9 GB peak) and sliding-window cache reuse (0.36 s to first token) are measured in the repo.
A benchmark for local models as coding agents: 20 tasks graded by the files left on disk, with a gate that tests every judge against known-good, known-bad and BOM-prefixed answers. Its main finding is that the harness is a variable: one model scored 12/12 in a lean engine and produced nothing inside a 66-tool harness.
A voice loop for Claude Code on macOS using Apple's speech recognizer, with stability-based endpointing and watchdogs for the recognizer's specific failure modes. A local Llama 3.2 3B on MLX decides whether each sentence was meant for the assistant.
A Mac render server that chains ACE-Step, a local Gemma, demucs, seed-vc and Whisper into a text-to-song pipeline, serving apps on the App Store and Google Play. It handles GPU contention, guards on model output, and karaoke timing that tolerates Whisper mishearing sung lyrics.
A one-file browser agent that drives the user's real browser over raw Chrome DevTools Protocol using a local model, with batching to cut model calls. Its prompt-injection guard lives in code rather than the prompt, so the model can't switch it off.
A proxy on Chrome's debugging port that lets several agents share one logged-in browser without seeing or closing each other's tabs. It filters CDP per connection, swallows focus-stealing and browser-quitting commands, and needs no client changes.
A terminal emulator written from scratch in Rust for Windows and macOS, with its own VT engine and native renderers. It rebuilds readable history from Claude Code's repaint-only screen by diffing screens, an approach built from captured bytes.