The posting
ABOUT AVRA
Avra is building relational foundation models for enterprise decision-making in Brazil.
Our work focuses on graph-native models for structured, high-stakes prediction problems: credit, fraud, growth, monitoring, and other decisions where entities cannot be understood in isolation. We model companies, people, and the relationships between them as evolving networks, then adapt those representations to customer-specific prediction tasks that plug into existing decisioning systems.
We work with internationally recognized research advisors, and we care about research that becomes useful in production.
THE ROLE
This is an applied scientist role with real modeling depth.
You will help evolve the thesis, architecture, and applications of Avra’s relational foundation models: how we train them, how we adapt them to specific tasks, and how they generalize across use cases.
Day to day, you’ll move between papers, code, experiments, and production constraints. The goal is not to try interesting ideas for their own sake. The goal is to find which ideas improve real downstream models under realistic deployment conditions.
We run a weekly research review. Strong papers matter; shipped models matter more.
WHAT YOU’LL WORK ON
- New approaches for relational foundation models over heterogeneous and temporal graphs
- GNNs, graph transformers, attention over relations, relative temporal encodings, and other architectures for structured entity networks
- Training objectives such as reconstruction, contrastive learning, generative modeling, supervised learning, and hybrid combinations
- Transfer from foundation representations to downstream tasks through fine-tuning, late fusion, distillation, calibration, and task-specific evaluation
- Rigorous evaluation: temporal validation, leakage checks, ablations, strong baselines, and error analysis
- Large-scale training infrastructure using Ray, including sampling, sharding, memory layout, distributed execution, and throughput optimization
- Performance-sensitive ML systems: data loading, graph sampling, memory efficiency, fused kernels, and training-loop bottlenecks
- Turning research ideas into reliable modeling components used in production
WHAT WE’RE LOOKING FOR
- 5+ years in applied ML research, research engineering, or equivalent high-level ML systems work
- Deep hands-on experience with PyTorch or a similar deep learning framework
- Ability to read current research, identify the core idea, and turn it into a controlled experiment within a week or two
- Experience with graph ML, recommender systems, ranking, time-series models, representation learning, or structured-data domains where strong tabular baselines are hard to beat
- Strong experimental discipline: baselines, ablations, temporal splits, leakage prevention, reproducibility, and honest error analysis
- Comfort with large datasets, distributed training, and the difference between a clean benchmark run and a pipeline that has to work every week
- Engineering judgment to build work that others can maintain
- Clear communication around model behavior, experimental results, and technical tradeoffs
YOU STAND OUT IF
- You have worked with heterogeneous or temporal graphs using PyG, DGL, custom graph tooling, or related systems
- You have used Ray for distributed training, data processing, or serving
- You have written Rust, C++, CUDA, Triton, or fused kernels, or worked seriously with JAX
- You have optimized graph sampling, memory usage, data loading, training loops, or distributed workloads
- You have shipped models into production and monitored how they behaved after deployment
- You have contributed to open-source ML infrastructure, published strong applied research, or built serious internal research systems
- You have worked in environments where the model only matters if it improves a real business metric
REQUIREMENTS
- Bachelor’s degree in a quantitative field: Computer Science, Mathematics, Statistics, Physics, Engineering, Economics, or similar
- Master’s or PhD is a plus, not a filter
- Strong written English
- Portuguese is useful, but not required
WHAT WE OFFER
- Competitive salary, equity, and open compensation bands
- Direct collaboration with founders, research leadership, and experienced AI advisors
- Research budget, paper incentives, and support for publishing when the work is strong and appropriate
- 100% remote work, with a São Paulo office available when you want it
- Flexible time off, national health plan, and extended parental leave
- High ownership over research directions that can become part of Avra’s core platform
If you want to help build foundation models for relational decision-making, not as a benchmark exercise but as infrastructure used by real enterprises on real economic networks, we’d like to meet you.
