Build the model, the system, and the company.

Own experiments across model architecture, training schedules, context, data mixtures, and scaling behavior.

• Run and analyze pretraining experiments.
• Improve capability per parameter and per token.
• Turn research ideas into reproducible training code.

Make training and inference faster, leaner, and more reliable on constrained hardware.

• Profile GPU and memory bottlenecks.
• Optimize precision, checkpointing, KV cache, and data movement.
• Build benchmarks that make performance regressions obvious.

Write and tune the low-level kernels behind the hottest parts of training and inference.

• Build CUDA kernels for attention, sampling, optimizer, and fused operations.
• Target consumer NVIDIA hardware first.
• Measure latency, occupancy, bandwidth, and memory behavior.

Build the pipelines that turn raw corpora into reliable, inspectable model-training data.

• Design dataset ingestion, filtering, deduplication, and sharding.
• Track provenance and data quality.
• Keep long-running jobs restartable and observable.

Build tests that tell the team what a model gained, what it lost, and why.

• Create capability and regression evaluations.
• Measure context use, instruction following, repetition, coding, and reasoning.
• Connect benchmark movement back to training decisions.

Turn model and systems work into polished tools people can actually use.

• Build fast interfaces around local models and research workflows.
• Work across frontend, backend, inference, and developer experience.
• Keep the product clear enough for users and useful enough for engineers.