Machine Learning Engineer

<div class="content-intro"><p><strong>About Neuralink:</strong></p> <p>We are creating devices that enable a bi-directional interface with the brain. These devices allow us to restore movement to the paralyzed, restore sight to the blind, and revolutionize how humans interact with their digital world.</p></div><p><iframe src="https://www.youtube.com/embed/XpxAMDrSJ7s" width="560" height="315"></iframe></p> <h3><strong>The opportunity</strong></h3> <p>In January 2024, a man paralyzed below the shoulders received a Neuralink implant. Within days he was playing chess and Civilization VI by imagining a cursor moving. That was our first product experience: computer control decoded from 1,024 electrodes in the brain’s motor cortex.</p> <p>Since then, 20+ people have been using the device, some for 17+ hours per day.</p> <p>We are now working on two problems nobody has solved.</p> <p><strong>Decoding the full virtual arm</strong>. This requires deciphering 29 degrees of freedom against millisecond timescale neural spikes. The published state of the art is four degrees of individuated finger control (Willsey et al., <em>Nature Medicine</em> 2025). We aim to fully decipher human intent of controlling anything that a human hand is capable of doing, and build a product that our users rely on for their independence every day.</p> <p><strong>Brain to voice</strong>. The frontier of real-time voice synthesis from intracortical signals is intelligible roughly half the time and carries only coarse pitch control (Wairagkar et al., <em>Nature</em>, 2025). We are working towards prosodic speech, in users’ own voices, streaming fast enough to hold a natural conversation.</p> <p>If you want to help us solve these problems, we want to hear from you.</p> <p> </p> <h3><strong>Why this is an interesting machine learning problem</strong></h3> <ul> <li><strong>Data.</strong> Train on >50,000 hours of neural data from clinical trial participants</li> <li><strong>Nonstationarity. </strong>Tackle the tough open problem that neural activity drifts and we need to solve decoding while minimizing user recalibration</li> <li><strong>Co-adaptation.</strong> The brain adapts to your decoder while your decoder adapts to the brain</li> <li><strong>Strict constraints. </strong>The brain implant operates on a tight power and the participant experience requires optimizing every millisecond of latency</li> <li><strong>Your eval is a person.</strong> Success is a human being capable of doing something today that they couldn't do yesterday</li> </ul> <p> </p> <h3><strong>What you'll work on</strong></h3> <ul> <li>Train models on neural spike data across sessions and participants so a new user’s decode works out of the box, and stays working through weeks of drift without supervised recalibration</li> <li>Design online adaptation that keeps the closed loop stable while both the model and the user are learning</li> <li>Compress a full hand-and-arm decoder to run with a tight latency budget on the user’s device</li> <li>Build the pre-training and post-training stack for voice synthesis: self-supervised representation learning on neural recordings, supervised fine-tuning on paired data, reinforcement learning with intelligibility and naturalness rewards, feeding a streaming vocoder that produces speech in real time</li> <li>Define what evidence justifies shipping a model update to someone’s brain-computer interface, then build the eval and rollout infrastructure to meet that bar</li> <li>Feed learnings back into the hardware platform. Our ASIC, thin-film arrays, and the rest of the brain implant are built in-house and follow your engineering gradient</li> </ul> <p> </p> <p><iframe src="https://www.youtube.com/embed/8TX-HhtDD2U" width="560" height="314"></iframe></p> <p> </p> <h3>You might be a good fit if you</h3> <ul> <li>Have designed, trained, and shipped real-time ML systems that people depend on</li> <li>Have strong sequence modeling instincts (e.g., in speech, robotics, reinforcement learning, time-series, sensor fusion)</li> <li>Prefer owning a problem end to end (data, model, deployment, eval) over one layer of a big stack</li> <li>Are deeply curious to understand and expand human consciousness</li> <li>Required: Bachelor’s degree (or higher) in Computer Science, a related field, or equivalent experience</li> </ul> <p> </p> <h3>Helpful (not required)</h3> <ul> <li>Familiarity with intracortical brain-computer interface decoding literature (e.g., handwriting and speech decoding, manifold alignment, unsupervised recalibration, cross-participant transfer)</li> <li>Experience with model compression, quantization, or inference on constrained hardware</li> <li>Experience in small-data or heavy-distribution-shift regimes</li> </ul> <p><em>Note: No neuroscience background is required. We value simple solutions built from first principles, and some of our best decoding work has come from people who have zero experience with neuroscience.</em></p> <p> </p> <h3>Our tech stack</h3> <ul> <li>Machine Learning: PyTorch, JAX</li> <li>Backend: Python, Rust, Swift, Bazel</li> <li>Frontend: Typescript, SwiftUI</li> </ul> <