Deep Learning Research Engineer Intern

DataRobot•Boston, WA

About The Position

DataRobot delivers AI that maximizes impact and minimizes business risk. Our platform and applications integrate into core business processes so teams can develop, deliver, and govern AI at scale. DataRobot empowers practitioners to deliver predictive and generative AI, and enables leaders to secure their AI assets. Organizations worldwide rely on DataRobot for AI that makes sense for their business — today and in the future. This effort is about building a multi-target joint probabilistic foundation model that can be used across industries to tackle some of the hardest real-world problems. The ambition goes beyond what tabular and time-series foundation models usually do: one model should support temporal forecasting, unordered tabular regression and classification, missing-data completion, and mixed-modality inputs and outputs while learning coherent joint structure across connected variables, rows, horizons, and scenarios. The primary objective is to create real value in serious use cases rather than optimize benchmark scores in isolation, although strong results on standard benchmarks are expected to follow closely. This role is designed as a research-engineering internship for someone operating at post-doctoral level, or very close to it, who wants to shape the architecture of a probabilistic foundation model. The center of gravity is the model itself: the encoders that read temporal, tabular, and mixed-modality inputs, the attention and sequence mechanisms that carry structure across variables, rows, and horizons, and the distributional output heads and decoding schemes that turn hidden states into coherent joint predictions. We are looking for someone who can reason precisely about what an architecture can and cannot represent, and who then implements, trains, and evaluates the resulting PyTorch models in a production-quality codebase. A background in stochastic modeling is a strong plus rather than a prerequisite: the model is trained on synthetic data drawn from stochastic dynamics, so a candidate who can also reason about stochastic differential equations and design synthetic data generators can contribute on both sides of the training loop.

Requirements

Strong PyTorch skills and hands-on experience building and training deep models.
Solid understanding of Transformers, attention variants, and long-context or state-space sequence architectures, including the tradeoffs between quality, memory, and latency.
Experience with probabilistic modeling in neural networks: distributional output heads, likelihood-based or proper-scoring-rule losses, mixture or flow models, or related uncertainty-aware learning setups.
Strong foundation in probability and statistics, enough to reason about densities and masses, dependence between variables, and calibration.
Ability to connect architectural ideas to working GPU-native implementations, controlled experiments, and diagnostics that show why a change helped.
Strong engineering habits: readable code, tests, reproducible experiments, and disciplined evaluation of model changes.
Ability to debug training instability, reason about why results changed, and iterate quickly from hypothesis to evidence.

Nice To Haves

Working knowledge of stochastic processes and stochastic differential equations: drift and diffusion, jumps, regime switching, mean reversion, heavy tails, and how such dynamics are simulated numerically.
Experience designing synthetic data generators or simulation-based training curricula, and an understanding of how the training distribution shapes what a model learns.
Depth in a domain with rich stochastic structure such as finance, energy, commodities, or a similarly quantitative field.
Familiarity with tabular or mixed-modality deep learning: categorical, ordinal, count, and bounded targets alongside continuous ones.

Responsibilities

Designing and improving the core model architecture: input encoders for temporal, unordered tabular, and mixed-modality data; attention factorizations across variables, rows, and horizons; the distributional output heads and decoding strategies that produce coherent joint samples.
Running controlled architecture studies, from ablations and scaling behavior to memory and throughput profiles, and turning the results into design decisions.
Building scalable PyTorch implementations that support larger input and output spaces, better throughput, and tighter memory budgets.
Studying how architectural choices, data-generation choices, and inference constraints interact to change benchmark quality and real-world usefulness.
For candidates with the stochastic-modeling background: extending the synthetic-data engine with richer stochastic dynamics, constraints, dependence structures, heavy tails, and regime behavior.
Turning research ideas into robust implementations and credible empirical results.