Avra · São Paulo ·
Builds and operates Avra's internal ML platform: CUDA kernels and distributed training for graph neural networks, columnar data formats (Lance/Arrow), experiment tracking, and a governed model registry so every model release is reproducible and auditable. Core stack: Python, PyTorch/Ray, CUDA, multi-node GPUs.
SimpliSafe · Boston, MA
Veeda AI · Toronto
arcade-ai · Presidio, CA
Aristocrat Leisure · Barcelona, Spain
At Avra, every technical IC is a Member of Technical Staff (MTS). The title doesn't put anyone in a silo: you own systems and outcomes, not steps in a function, and you keep building depth in your area.
In this role, you'll join our ML Systems team, which owns Avra's ML core and the governance of every model we ship. Research produces candidate models and evidence; you build the reliable path from data and training to a governed, reproducible release that can run in our cloud or in any customer environment. ML Systems is an internal platform: its users are our researchers and platform engineers, and its success is measured by the leverage it creates for them.
Build CUDA kernels and compute primitives for training and serving graph neural networks (GNNs).
Evolve Monad, our sampler and distributed-training library, including neighbor sampling and training performance.
Specify our binary data formats (Lance, Arrow, CSR/CSC), and own materializations and feature backfills for training and evaluation.
Define data contracts and consumption requirements with the teams that build our customer and proprietary datasets.
Build and operate experiment tracking, checkpoints, and evaluation infrastructure, with reproducibility by default.
Own the model registry, lineage, versioning, and compatibility across models, embeddings, and downstream models.
Define and run release gates, so every model running in production, batch, or on-premise maps to a governed release.
Make it possible to audit exactly which data, code, configuration, and evidence produced each release.
Time-to-experiment: how quickly a researcher goes from a hypothesis to materialized data, compute, and tracking.
Time-to-governed-release: how quickly a validated candidate becomes an authorized release.
Training throughput per GPU on our foundation model training runs.
100% of production models with complete release records and lineage — no ad hoc models in any environment.
Every release reproducible from its registered data, code, and configuration.
Strong systems engineering skills and production-quality Python.
Experience with distributed training (e.g., Ray, PyTorch distributed) and multi-node GPU workloads.
Experience with columnar data formats and large-scale data materialization.
Familiarity with ML lifecycle tooling: experiment tracking, model registries, evaluation, and reproducibility.
A product mindset: you treat an internal platform as a product with real users. You don't need to be a data scientist.
CUDA kernel development or GPU performance optimization.
Graph neural networks or graph sampling at scale.
Lance, Arrow, or other columnar/indexed storage formats.
Multi-cloud GPU compute (e.g., SkyPilot).
Model governance or audit requirements in financial services or other regulated environments.
Apple · Cupertino