Nina Kowalski

I live in PyTorch: custom modules, distributed training, and turning paper ideas into reproducible training stacks—then handing off traced/quantized builds to platform teams.

• Shipped domain-adaptive vision transformers with PyTorch Lightning + Weights & Biases sweeps
• Built LoRA fine-tuning harness for 7B-class LLMs with eval harnesses and rollback-safe checkpoints

1. Staff ML Engineer

   Polar Flux Labs · 2020 — Present

   Staff ML engineer at Polar Flux Labs: PyTorch-centric research loops for multimodal models; Lightning + W&B sweeps with regression tests on CUDA upgrades. LoRA/QLoRA harness for 7B-class LLMs with eval gates, rollback-safe checkpoints, and human review queues. TorchBench-style ablations on diffusion schedulers; CI integration for kernel upgrades.