Torin EtheridgePortfolio / 2026Software engineering, designIndex / 01–06

TorinEtheridge

00 / Introduction

Software engineer building precise, design-led products.

Selected work

  1. 01AlpacaNative language model inference for Apple Silicon, in Swift and Metal, with no llama.cpp, MLX or Core ML underneath.Swift, MetalInference engine2026
  2. 02RelayA durable job queue in Rust, with leases and fencing tokens.Rust, SQLiteDistributed systems2026
  3. 03VecStreamVector search with HNSW written from scratch.Python, NumPySearch + storage2025–26
  4. 04PrismGPU image processing for Swift, built on an explicit render graph and custom Metal kernels.Swift, MetalGraphics engine2026
  5. 05PaletColors from the world around you, extracted on the device.SwiftiOS + product2026
  6. 06ObjectsIndustrial design and physical products.W01, L01Objects + design2026

Plates

Fig. 01–05
200400600327305F16549430Q8_0671503Q4_0DECODE, TOKENS PER SECONDALPACALLAMA.CPPSMOLLM2-135M / M5 / CONTEXT 16

Fig. 01 / Alpaca, Decode throughput against llama.cpp549 tok/s · 1.28× llama.cpp decode · 72/72 greedy tokens match

SUBMITDELAYED, DUECLAIMACKLEASE LOSTFAILRETRY DUEEXHAUSTEDREPLAYPENDINGREADYLEASEDSUCCEEDEDRETRYINGDEADFENCING TOKEN ON EVERY LEASE

Fig. 02 / Relay, Job lifecycle9,792/s · 270 µs p99 · +129% after profiling

0.80.91.00.51.01.5EF 10 0.768EF 20 0.908EF 40 0.990EF 80 0.996EF 160 1.000RECALL@10P50 LATENCY, MS

Fig. 03 / VecStream, Recall against latency0.99 recall@10 · 1.1 ms p50 · 44 tests

BLURREDORIGINALINPUTEXPOSUREBLUR HBLUR VSHARPENOUTPUTRENDER GRAPH / VALIDATED DAGSHARPEN READS THE ORIGINAL AND THE BLURRED IMAGE

Fig. 04 / Prism, Render graph, sharpen pipeline3.3 ms 4K blur · −57% after profiling · 182 tests

Palet Camera capture screen.Palet Saved palettes screen.Palet Palette detail screen.

Fig. 05 / Palet, capture, library, detailSwift · iOS · local processing

Elsewhere