AI Researcher / Engineer / Intern

Hi, I'm Brian, co-founder of Egra. This posting is one role with wide possibilities: full-time researchers, full-time engineers, interns, college kids who've been shipping since they were 16. The bar is the same. You'll have complete ownership over your work from day one. No lengthy onboarding, waiting for permission, or navigating layers of approval. Real problems, real compute, real autonomy. The work you do in your first month will be in the product. What you'd be doing We're training models on the highest-bandwidth data anyone has tried to model at scale. Biosignals, eye-tracking, pupillometry, video of human faces, audio, the content humans are reacting to, downstream behavior. A non-exhaustive list of projects on the table: - Designing self-supervised pretraining objectives on multimodal physiological + content data - Stress-testing recent multimodal / signal foundation model papers to understand exactly where and why they break under distribution shift - Building evaluation protocols that distinguish real progress from leaky-benchmark noise - Shipping the internal research tooling (experiment tracking, dataset versioning, agentic eval pipelines) that lets research compound instead of repeat - Closing the loop between offline model results and the live product - Writing the internal research memos that become the shared knowledge base: "why model X fails on dataset Y," "what we tried and why it didn't work" You should be able to take any one of these and have something running end-to-end in a week. We don't separate "research" and "engineering". The people we want are both. Where this is going We're building toward a future where AI is actually optimized for what it does to the human on the other side of the screen. The current signal AI is trained on is a shadow of what matters. Research culture A few strong opinions: Bitter Lesson by default. We're skeptical of hand-engineered features and domain-specific heuristics. We'd rather let models discover structure than force-feed it. Reproducibility over vibes. If we can't tell you which preprocessing version produced a result, we don't trust the result. Failed experiments are documentation, not waste. Write up what doesn't work with the same care as what does. AI-augmented everything. Claude Code, Cursor, internal eval agents, generated tooling. Anything one person can do with AI tools is owned by one person. Who we're looking for The shape we want, in priority order: - You can take a vague research direction and ship something concrete within a week. - You have strong opinions about what makes representations actually generalize, and they're informed by experiments you ran yourself. - You're comfortable with heterogeneous, multimodal data and you have a toolkit for making it useful. - You ship fast with AI coding tools. Codex, Claude Code, agents. - You have taste. You can look at a benchmark and tell us why it's leaky. You can look at a model architecture and tell us what it can't possibly learn. Background doesn't matter much. PhD, no PhD, dropped out, finishing undergrad — all fine. We've hired interns who out-shipped postdocs and we'll do it again. You should NOT apply if: - Your value proposition is "I know EEG" or "I have a neuroscience background." This is an ML role, not a neuro role. We are not hiring for domain expertise in any signal modality. We're hiring for general modeling ability on hard data. - You need a clear roadmap or a manager to do your best work. - You're more interested in neuroscience theory than building systems that work in production. - You rely heavily on hand-crafted features, classical signal-processing pipelines, or domain-specific engineering as your main contribution. - You want to publish first and ship second. Interview process Three conversations, total ~90 minutes: 30-min intro. We tell you what we're building, you tell us what you've built. Casual, no prep. 30-min technical jam. We want to see how you think about problems, approach t