Technical white papers spanning work from 2018 to today, plus ongoing academic research in mechanistic interpretability and on-device AI. Grouped by the year each project was actually built, not when we got around to documenting it.
An NCCL‑equivalent for Apple Silicon clusters —
auto‑tuned topology (ring / tree / hierarchical) based on
measured Thunderbolt and Ethernet link bandwidth, in‑flight
lossy compression on slow links, and a C/Swift API that mirrors
the ncclComm_t idiom. Fills the role NCCL plays on CUDA
hardware for Mac clusters.
A Triton compiler backend targeting Apple Metal, with a Swift core
and a thin Python shim. Existing @triton.jit source runs
unmodified on Mac GPUs — matmul at 75–76% of Apple’s
own MPS, FlashAttention‑2 beating the MPS composite at
s ≤ 1024 in both directions. Breaks the portable‑kernel
moat Triton gave CUDA‑only targets.
An ML‑native kernel profiler for Metal — what Nsight Compute
is for CUDA. Computes arithmetic intensity analytically from ML kernel
shapes and scores against ceilings measured on the local chip (not
spec‑sheet peaks). Per‑encoder GPU timestamps via
MTLCounterSampleBuffer, static occupancy from
MTLComputePipelineState, and a kernel differ.
A native iPadOS app that turns a tablet and RTL‑SDR dongle into a full‑spectrum scanner and automatic emitter classifier — 24 MHz to 1.766 GHz, with a Metal waterfall display and DSP pipeline in Accelerate/vDSP.
Read PDF ↗A production‑grade Swift 6 inference server on Hummingbird 2 and MLX Swift — OpenAI‑compatible API, prefix‑trie KV cache, deadline‑bounded batch assembler, and LRU model pool. No Python, no Docker.
Read PDF ↗GPU‑accelerated TSDF volumetric reconstruction for watertight 3D scanning on consumer LiDAR hardware — real‑time depth‑frame integration via Metal compute, Marching Cubes surface extraction, OBJ/STL export ready for any slicer.
Read PDF ↗On‑device neural TTS for academic reading on iPad — Kokoro 82M (StyleTTS2) reimplemented in MLX Swift, with a full G2P frontend, PLBERT prosody encoder, and iSTFT vocoder. Zero network calls at inference time.
Read PDF ↗A native Swift canvas for visual data science and compound model orchestration on Apple Silicon.
Read PDF ↗A native macOS and iPadOS app that brings LLM quantization — GGUF and MLX pipelines — out of the terminal and into a direct‑manipulation interface. Inspect architecture, choose a format, watch live progress, get before/after benchmarks.
Read PDF ↗A federated protocol for privacy‑tiered, measured‑accountability AI compute across wide‑area networks.
Read PDF ↗A native iPad control plane for distributed AI inference clusters built on exo.
Read PDF ↗A universal platform for neural network interpretability, attribution, and surgical intervention.
Read PDF ↗An MLX‑native Swift reimplementation of the Python scientific data science stack.
Read PDF ↗On‑device few‑shot annotation and dataset organization for Apple platforms. Originally built 2020, actively maintained since.
Read PDF ↗A universal platform for on‑device machine learning training, inference, and model management — letting anyone build Core ML models from their own data. Our earliest work.
Read PDF ↗Mechanistic interpretability, sparse autoencoders, and on-device AI efficiency — academic papers in active review at ICLR and NeurIPS. Preprints available now.
Circuit‑level mechanisms identified in small models generalize to larger models with additional redundancy structure. Targeted path‑patching combined with sparse decomposition on Apple Silicon — IOI circuits measured on Llama and Pythia.
Abstract & PDF ↗TopK SAE training dynamics across Llama, Mistral, and Qwen: dictionary collapse, dense‑feature degeneracy, and held‑out reconstruction. Evaluation metric choice determines which architecture appears superior — a hidden confound in prior work.
Abstract & PDF ↗Activation streaming across two nodes via Thunderbolt: bit‑exact activation splits at 1.25 GB/s with sub‑millisecond scheduling overhead. A graph‑based attribution method propagating credit across the full residual stream for safety‑relevant behaviors.
Abstract & PDF ↗A controlled audit of 4‑bit MLX quantization cost for code generation on Apple Silicon. Measures perplexity, HumanEval/EvalPlus pass rates, and tokens/sec across quantization levels under a reproducible harness.
Abstract & PDF ↗The pre‑lab years. More than twenty apps shipped since 2011, across domains with nothing in common — where the range came from.
More papers are in the works. Open a channel if you want to talk about any of them.