XVI LABS

The role

You'll train and post-train the models at the core of XVI Labs: models that learn from records of real decisions and predict what a specific person will do next, and why. We're early, so you'll help choose architectures, run the experiments that show what works, and take the winners to scale.

What you'll do

• Pretrain and post-train models on large-scale market decision data, from data loading through evaluation
• Fine-tune open-source models to predict an individual's next decision from the context they had before making it
• Design and run controlled experiments: baselines, ablations, scaling runs and go/no-go tests written down before the run
• Build evaluations that show whether a model has learned real decision behavior or only surface patterns
• Own training infrastructure across multi-GPU runs, from hyperparameter sweeps to large training jobs

What we're looking for

• Hands-on experience training large models and post-training models
• Strong PyTorch and fluency with modern training stacks: distributed training, mixed precision, checkpointing
• A deep understanding of sequence architectures, from transformers to state-space models
• Rigor in experiment design: baselines first, and loss curves trusted only as far as the evals back them up
• Comfort taking an open research question all the way to a working result

Strong candidates also have

• Experience with time-series, financial or market microstructure data
• Published ML research or substantial open-source work
• Experience with reinforcement learning or preference-based post-training
• A background in quantitative trading, decision science or behavioral economics
• Experience as a founder or early engineer at a zero-to-one company

The role

You'll build the systems that turn raw records of human decisions into training data, and the systems that put our models in front of customers. You'll work across data, infrastructure and product, side by side with the research team.

What you'll do

• Build ingestion pipelines for large decision datasets, starting with financial and prediction market data and expanding to new domains
• Design storage and query layers for very large volumes of timestamped events
• Build data quality tooling: deduplication, actor labeling, and leakage checks between training and evaluation data
• Ship the interfaces customers use to query our models
• Own reliability, testing and deployment across the stack

What we're looking for

• Strong software engineering fundamentals and production experience in Python
• Experience building data pipelines that stay correct at scale, with care for time ordering and edge cases
• Comfort with cloud infrastructure, containers and CI
• Experience in machine learning (preferred)
• Ownership: you ship, measure, and fix what you shipped

Strong candidates also have

• Experience with financial market data, blockchain data or other high-volume event streams
• Familiarity with columnar tools such as Parquet, DuckDB, Polars or ClickHouse
• Experience supporting ML training: data loaders, GPU infrastructure, experiment tracking
• Experience as a founder or early engineer at a zero-to-one company