Member of Technical Staff, Research — Relational Foundation Models

About Avra

Foundation models learned to read text, images, and code. The data that runs the economy hasn't had its moment yet. It lives in relational databases: customers, accounts, transactions, and claims, linked by keys and moving through time. Every credit approval, fraud block, and retention offer is a prediction over that data. Today each one is still a bespoke model, rebuilt by hand for every task, schema, and company.

We think that ends with a foundation model for relational data: one model that reads any database, understands its schema from the names of its tables and columns, and predicts from labelled examples shown to it in context, the way language models learned to follow a prompt. That means no training run per task and no feature-engineering project per customer.

Avra exists to build that model and the platform that turns its predictions into decisions running in production, in our cloud or inside an enterprise's own infrastructure. We're a frontier lab with its home in São Paulo, backed by Sequoia and Andreessen Horowitz, and we ship with enterprise design partners in banking, payments, telecom, and commerce, including Mercado Livre, Inter, PicPay, Banco BV, and Vivo.

About the team

We are a small group of researchers, engineers, and company builders, with our home in São Paulo and a team spread across San Francisco, Latin America, and Europe. Everyone technical at Avra is a Member of Technical Staff. We don't separate "researchers" from "engineers" before the problem is understood, so the people who design an experiment also write the code that runs it and own what happens next. We hire at all levels. The scope you take on comes from what you've shown, not from a title.

Your mission

Make one pretrained model predict well on relational databases it has never seen, from the examples in its context.

What you'll work on

• In-context learning over relational data. How the model uses labelled examples at inference time, what makes a good context (which rows, how much history, which neighbours), and why it fails when it fails.

• Pretraining from scratch. Architectures, objectives, and scaling for transformers over multi-table, temporal data, on multi-GPU H100 runs.

• Pretraining data. Which corpora, real and synthetic, teach a model to generalise across schemas, domains, and languages.

• Time and leakage. Predictions that only see what was knowable at decision time. That's the difference between a benchmark number and a model a bank can trust.

• Evaluation. Relational and tabular benchmark suites, strong baselines (gradient boosting, tabular foundation models, task-specific relational models), and protocols that keep comparisons honest.

• Beyond the context window. When fine-tuning or post-training on a customer's database beats in-context prediction, and by how much.

What we hope you'll bring

• Excellent fundamentals and excellent implementation skills. You train models in PyTorch yourself, including multi-GPU runs you've debugged.

• Research taste: you can read a paper, find the idea that matters, and design the cheapest experiment that could prove it wrong

• Experimental discipline: temporal splits, ablations, multiple seeds, strong baselines, and honest error analysis

• Experience with structured data, such as tabular or relational learning, recommender systems, time series, or any domain where gradient boosting is hard to beat

• Clear written English

Bonus points if you have

• Worked on tabular or relational foundation models, in-context learning, or meta-learning

• Pretrained transformers at scale, or studied their scaling behaviour

• Written Rust, C++, CUDA, or Triton, or made a data loader or sampler substantially faster

• Shipped a model to production and lived with how it behaved afterwards

• A record of strong research, whether in papers or in open-source work

We welcome academic researchers and researchers with unconventional backgrounds. A degree is not a requirement.

Logistics

• Location: Remote, from anywhere. The team works across San Francisco, Brazil, Uruguay, and Europe, and we have an office in São Paulo.

• Compensation: Competitive salary and equity, depending on background and level.

• Compute: H100-class GPUs for pretraining-scale experiments, plus a research budget.

• Benefits: Flexible time off, health coverage, extended parental leave, and support for conferences and publishing.

• Language: We work and interview in English.

This is an evergreen role. We keep it open, review every application, and reach out as positions open. Please wait six months before you apply again.