The systems layer under machine learning: training libraries, checkpoints, agent evaluation and LLM observability.
I've been building software since 2018, most of it in Rust: async services, FFI, memory allocators and protocol code that has to stay correct under load.
Today I work where that background matters most in AI: the infrastructure that models and agents run on. That means numerical correctness in training code, checkpoints that survive a crash, evaluation scores you can trust, and traces that tell you why an agent did what it did.
Models are only as reliable as the systems that train, evaluate and serve them.
| Area | What I focus on |
|---|---|
| Training library internals | Optimizer math (RMSprop variance, epsilon placement), weight initialization (Kaiming/Xavier), RNN input validation, numerically stable edge cases |
| Checkpointing & data integrity | Atomic checkpoint writes, batched tensor serialization, verified dataset downloads |
| Agent evaluation | Scoring and weighting of eval dimensions |
| LLM observability | OpenTelemetry spans for agent and tool events, error capture in LangChain tracing, token-usage parsing |
| Forecasting & anomaly detection | Exogenous-feature support in time-series foundation-model pipelines |
| Systems foundations | Tokio, cancellation safety, backpressure, zero-copy I/O, FFI and unsafe audits |
All links go to my own pull requests.
Merged
| Project | Contribution |
|---|---|
| FastNN: neural-network library in Rust + CUDA | Reject zero-length sequences in LSTM/GRU with a clear message |
| Verify MNIST download integrity before using a cached file | |
| jamjet: safety layer for AI agents | Fix DimensionScores.composite() weight normalization |
| Add OpenTelemetry span attributes for coordinator and agent-tool events |
In review
| Project | Contribution |
|---|---|
| FastNN | Fix centered RMSprop: clamp variance, move eps outside sqrt |
| Atomic checkpoint writes · Batch tensor I/O in checkpoint serialization | |
Assert fan_in > 0 in Kaiming/Xavier init |
|
| foundationforecast: time-series foundation models | Accept X_df and futr_exog_list in detect_anomalies |
| lemma: LLM tracing | Record handled LangChain tool errors as span errors |
| oxtrace: LLM usage tracing | Parse usage from gzip-encoded responses |
| runanywhere-sdks: on-device AI SDKs | Typecheck core test files |
Merged contributions to py-libp2p, papi-console (Polkadot), circom, beancount-language-server, perun-ckb-backend, uor-addr and metal2vk.
Own repositories: smalloc (memory allocator for Solana programs), Dbgeng-sys (Rust FFI bindings for the Windows debugger engine), Solana-Credit-Score, AsyncCSVX.
01. Measure, then optimize. Throughput, tail latency and allocations, not guesses.
02. Make failure boring. Writes are atomic, inputs are validated at the boundary, and errors say what to do next.
03. Trust numbers only when they're tested. An eval score or a loss curve is only as good as the code that computes it.
04. Small, reviewable changes. Every PR above is scoped to one problem and easy to review.
Open to remote roles and contracts with AI/ML startups: ML infrastructure, inference, evaluation and Rust performance work.