Senior Rust engineer/AI/ML infrastructure
I build whatAI runs on.
Training libraries. Checkpoints. Evals. Traces.
I'm Suzuki Taro, and I make the systems under your models correct.
01 / 05
Training
Optimizers, initialization and RNN layers in Rust.
Depth
Not just another model wrapper
The system
comes first.
A silent numerical bug or a corrupted checkpoint can cost days of GPU time. I start with the question every infrastructure team should ask: what has to stay true when things fail?
My background is high-performance Rust: async services, FFI, memory allocators and distributed protocols that have to stay correct under load. I bring that discipline to the code that trains, evaluates and observes AI.
What I do
04 areas
- 01
Training internals
↗Optimizer math, weight initialization, RNN input validation and the numerical edge cases that quietly corrupt a run.
OPTIMIZERS · INIT · LSTM/GRU
- 02
Checkpoints & data
↗Atomic checkpoint writes, batched tensor serialization and verified dataset downloads, so a crash never costs a day of compute.
ATOMIC I/O · TENSORS · INTEGRITY
- 03
Agent evaluation
↗Composite scores that weight dimensions correctly, so the number on the dashboard means what you think it means.
EVALS · SCORING · SAFETY
- 04
LLM observability
↗OpenTelemetry spans for agents and tools, error capture in LangChain tracing and token-usage parsing.
OTEL · TRACING · TOKENS
Open source log
Shipped.
- MergedFastNN #78Reject zero-length sequences in LSTM/GRU
- MergedFastNN #72Verify MNIST download integrity before using a cached file
- Mergedjamjet #176Fix composite eval-score weight normalization
- Mergedjamjet #197OpenTelemetry span attributes for agent-tool events
- In reviewFastNN #73Fix centered RMSprop: clamp variance, move eps outside sqrt
- In reviewlemma #137Record handled LangChain tool errors as span errors
Technologies: Rust, Tokio, Python, TypeScript, OpenTelemetry, FFI, Distributed systems
Open to remote roles · US time zones