Mustard is a multi-instance watch-party platform: clients connect over WebSockets and use NTP-style clock estimation with predictive drift correction, and updates go through atomic Redis Lua scripts. The 48 ms P95 is a specific claim: steady-state player-reported drift, measured across three Chrome clients at approximately 300 ms RTT during a 240-second test. It is not a physical audio-output measurement.

Protocol correctness was a separate question. TLA+/TLC model checking exposed a stale-epoch bug after store resets, leading to an ordered-epoch fix, and idempotency keys plus atomic updates keep control commands from being applied twice within the deduplication window.

I also built Go and Rust relays to test protocol conformance and runtime costs. They are study implementations, not the production backend. In a local 10,000-connection test, the Rust relay used 14.5 KB of memory per connection versus 40 KB for Go; the extreme-tail latency comparison was inconclusive.

Live site · Source · Sync design · TLA+ findings · Go/Rust study

These LayerNorm/RMSNorm kernels combine fused residual-add, FP8 outputs and deterministic backward reductions, and come as drop-in PyTorch modules that work under torch.compile without graph breaks. Parameter gradients use fixed-order reductions rather than atomic accumulation, which is where the determinism comes from.

For dynamic RMSNorm-to-FP8, I measured 4.9–7.2× over the eager PyTorch composite and 1.04–1.76× over the compiled composite, across tested FP16 shapes on an NVIDIA A100. Those are kernel-time comparisons, not end-to-end model speedups. For a fuller picture, the repository also documents H100 results and configurations where PyTorch wins.

Source · A100 results · Methodology

Rather than guess at a training configuration, the tuner benchmarks precision, compile mode, memory format and fused-optimizer configurations on the target model and batch. Search runs under a budget, configurations are cached for reuse, and reports show trial results, compilation cost and estimated break-even time.

I measured 2.7–6.7× training-step speedups after tuning versus the PyTorch-default FP32 eager baseline across ResNet-50, ResNet-18 and a six-layer Transformer on an NVIDIA A100, with the tuner using torch.compile alongside the other optimizations. The baseline matters, though: on the Transformer, the gain is 2.55× against an eager baseline with TF32 matmuls enabled, rather than 6.65× against the defaults. Search costs and numerical trade-offs are documented as well.

PyPI · Source · Benchmark artifacts

Spark is a Chrome extension I built and published for capturing and searching conversation history. A React/TypeScript dashboard lets you browse saved conversations, FastAPI handles backend requests, and ChromaDB stores the embeddings behind semantic search.

Source · Chrome Web Store