About the Role
This posting is for a role with wide possibilities: full-time researchers, full-time engineers, and interns. The bar is the same for all. You will have complete ownership over your work from day one, with no lengthy onboarding or layers of approval. The work you do in your first month will be integrated into the product.
What you'd be doing
You will be training models on high-bandwidth data, including biosignals, eye-tracking, pupillometry, video of human faces, audio, and the content humans are reacting to. Projects include, but are not limited to:
- Designing self-supervised pretraining objectives on multimodal physiological + content data.
- Stress-testing recent multimodal / signal foundation model papers to understand where and why they fail under distribution shift.
- Building evaluation protocols to distinguish real progress from leaky-benchmark noise.
- Shipping internal research tooling (experiment tracking, dataset versioning, agentic eval pipelines) to facilitate research progress.
- Closing the loop between offline model results and the live product.
- Writing internal research memos to build a shared knowledge base, documenting both successes and failures.
You should be able to take any of these projects and have something running end-to-end within a week. We do not separate "research" and "engineering"; the ideal candidate embodies both.
Research Culture
Our research culture is built on a few strong opinions:
- Bitter Lesson by default: We are skeptical of hand-engineered features and prefer models to discover structure.
- Reproducibility over vibes: Results must be traceable to specific preprocessing versions to be trusted.
- Failed experiments are documentation, not waste: Document what doesn't work with the same care as what does.
- AI-augmented everything: Leverage tools like Claude Code, Cursor, and internal eval agents for maximum efficiency.
Who we're looking for
In priority order, we seek individuals who:
- Can take a vague research direction and ship something concrete within a week.
- Have strong, experimentally-informed opinions on what makes representations generalize.
- Are comfortable with heterogeneous, multimodal data and possess a toolkit for making it useful.
- Can ship fast using AI coding tools such as Codex, Claude Code, and agents.
- Possess strong taste, able to identify leaky benchmarks and architectural limitations.
Background (PhD, no PhD, dropped out, finishing undergrad) is not a primary factor. We have a history of hiring interns who outperform postdocs.
You should NOT apply if:
- Your value proposition is domain expertise in specific signal modalities (e.g., "I know EEG" or "I have a neuroscience background"). This is an ML role, not a neuro role.
- You require a clear roadmap or a manager to perform your best work.
- You are more interested in neuroscience theory than in building production-ready systems.
- You heavily rely on hand-crafted features, classical signal-processing pipelines, or domain-specific engineering.
- You prioritize publishing over shipping.
Benefits
- Competitive salary and meaningful equity (for full-time roles); top-of-market intern stipend.
- Platinum-tier health insurance.
- Uncapped compute access and AI tooling budget.
- Full research autonomy with no bureaucracy or review committees.
- Relocation and visa support; in-person work is strongly preferred, with flexibility for edge cases.