Want to know if this job is worth applying to?
São Paulo, São Paulo, Brasil
Remote only
Avra is a frontier AI lab building relational foundation models from the ground up (no third-party wrappers).
We model how businesses and individuals are connected and interact over time. Put simply: we bring connections into the math.
Leading enterprises in LatAm use Avra's models to power key decisions in credit, growth, and fraud. Our customers include the largest banks and marketplaces in Latin America, such as Mercado Libre, Telefonica, Banco Inter and Getnet.
Avra was founded in 2024 by Viviane Meister and Bruno Alano (ex-OpenAI). Our investors include NXTP, Clocktower Ventures, as well as the Sequoia and a16z scouts. We are accelerated by NVIDIA, AWS, and Google.
We are a team of builders. Join us.
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 the Platform team to own where our models execute. Customers consume our models through large batches of millions of records and through real-time APIs, and they make business decisions on every response. You'll run governed model releases reliably and efficiently — in our cloud and on customer-hosted Kubernetes — and make inference fast, predictable, and cheap enough to serve both enterprise and mid-market customers.
Evolve Sophos, our online and batch inference runtime, built on Kubernetes.
Run large batch inference on ephemeral jobs, with multi-dimensional admission control (CPU, memory, GPU).
Build and extend the controller and its Kubernetes custom resources.
Optimize each model's inference engine and feature processing.
Serve graphs and data efficiently.
Own execution of training, post-training, and fine-tuning jobs, in our cloud and in customer dataplanes / on-premisse cloud.
Drive autoscaling, GPU serving, performance, and cost optimization, with telemetry for every model we run.
99.9% serving availability.
p95/p99 latency for online inference and throughput for batch.
Cost per prediction and per training job.
GPU utilization: paid capacity versus capacity actually used.
Training and batch jobs that finish on time and succeed without manual retries.
Experience running model serving or large-scale batch compute on Kubernetes.
Experience building Kubernetes controllers or operators.
Skill at profiling and optimizing data-heavy Python pipelines.
A clear sense of cost: you treat compute efficiency as a product feature.
Production-quality code and reviews, and a willingness to operate what you build.
Ray, Ray Serve, or KubeRay in production.
Admission-control systems.
GPU serving and performance optimization.
Arrow, Parquet, Lance, or other columnar formats.
Shipping software to customer-hosted Kubernetes.
GCP/AWS and GKE/EKS, and financial services or regulated environments.
