Research Engineer, ML Infrastructure

Cognition•San Francisco, CA

9d•Onsite

About The Position

We are an applied AI lab building end-to-end software agents. We're the makers of Devin, the first AI software engineer. Our team is extremely talent-dense. Among our founding team, we have world-class competitive programmers, former founders, and leaders from companies at the cutting edge of AI including Scale AI, Palantir, Cursor, Waymo, Tesla, Lunchclub, Modal, Google DeepMind, and Nuro. Building Devin is just the first step—our hardest challenges still lie ahead. If you’re excited to solve some of the world’s biggest problems and build AI that can reason on real-world tasks, apply to join us. Role Mission Research moves at the speed of the infrastructure underneath it. Every training run, evaluation loop, and experimental iteration depends on systems that are fast, reliable, and built to scale. This role exists to make sure nothing in the stack becomes the bottleneck that slows down the frontier. You will own the core systems that researchers depend on daily: distributed training infrastructure, experiment orchestration, data pipelines, and the tooling that turns raw compute into usable research velocity. This is not a support role. You will work directly alongside researchers, understand the science deeply enough to anticipate what they need next, and build systems that hold up under the pressure of training jobs running across thousands of GPUs. We don't distinguish between research and engineering; the best infrastructure engineers here are also the ones who understand why the research works.

Requirements

Deep experience building and operating distributed training systems for large models; comfortable owning infrastructure end to end from the cluster level down to the training loop
Strong systems engineering fundamentals: distributed systems, networking, storage, and the ability to reason about performance across the full hardware-software stack
Proficiency in Python and C++; experience with PyTorch or equivalent deep learning frameworks at a systems level, not just API usage
Hands-on experience with GPU performance profiling, memory optimization, and compute efficiency; able to diagnose why a training run is underperforming and fix it
Experience implementing or optimizing parallelism strategies (data, tensor, pipeline, sequence) for large model training
Track record of building tooling and abstractions that meaningfully accelerate research workflows
Strong debugging instincts across complex, distributed systems where failures are non-deterministic and hard to reproduce
Enough ML knowledge to engage substantively with researchers: understand what they are training, why the architecture choices matter, and what the infrastructure needs to support

Nice To Haves

A PhD is one signal among many.

Responsibilities

Build and own the systems that run large-scale training jobs reliably across GPU clusters. This includes job launchers, checkpointing and recovery, fault tolerance, and the monitoring that keeps researchers informed and unblocked.
Own the infrastructure that runs hundreds of thousands of concurrent coding agent rollouts in VM sandboxes, from high-fidelity environment design to the distributed systems that hold up at our largest RL training scales.
Profile and improve training throughput end to end. Identify bottlenecks across data loading, communication overhead, memory utilization, and compute efficiency. Implement solutions that meaningfully improve step time and MFU at scale.
Design and maintain the systems researchers use to launch, track, and analyze experiments. Reduce friction in the research loop so that more time is spent on ideas and less on waiting.
Build high-throughput, reliable data pipelines for training and evaluation. Ensure data quality, reproducibility, and efficiency at the scale our training runs demand.
Diagnose and resolve training failures across GPUs, networking, numerics, and data. Maintain detailed understanding of failure modes and build systems that fail gracefully and recover fast.
Implement and optimize parallelism strategies: data, tensor, pipeline, and sequence parallelism. Understand the tradeoffs deeply and apply them to get the most out of available hardware.
Anticipate what the research team will need next and build it before it becomes a constraint. The best infrastructure engineers here are proactive, not reactive.