ML Infrastructure, self-hosted LLMs.

Replacing fragmented ML tooling with a minimalist provider-agnostic control plane that runs ML workloads across multiple GPU providers. Enabling 200 ML practitioners to run self-hosted models without touching infrastructure.

Efficient on-prem batch/stream data pipelines.

Migrated flight telemetry data platform from AWS/GCP to an on-prem instance, improving data ingestion performance by 3x across 5 data formats. Streamlined new customer onboarding by building reusable batch/stream pipelines.

Got Zelensky's voice to the world.

Built an AI voice dubbing platform in weeks, partnered with the Office of the President of Ukraine, and shipped daily YouTube updates in multiple languages. Pivoted into information retrieval from audio, open-sourced as Platogram.

Scaled PyTorch's audio library through hypergrowth.

Inherited torchaudio team and helped scale to 1M+ monthly downloads. Supported the team through pandemic chaos. As an IC, improved XLM-R-XL fine-tuning success rates by integrating PyTorch Lightning with Meta's infrastructure.

Automated ML workflows and grew a world-class team.

Co-designed a Kubernetes-based MLOps platform that lets researchers focus on models, not infrastructure. Recruited two Python Core developers. Shipped MVP, established engineering culture, and supported ML consulting.