Research Intern

Egra•New York, NY

About The Position

Hi, I'm Brian, Co-Founder of Egra. We just raised $5.5M to build foundation models for brain signals, and we're looking for a research intern to join our team for the summer (or longer). This isn't a "shadow someone and take notes" internship. You'll work directly alongside the founding team on real research problems from day one. You'll run your own experiments, contribute to our shared knowledge base, and leave with work you're genuinely proud of. If you want high agency, hard problems, and the chance to shape early-stage research at a funded startup, this is that opportunity. What you'd be doing EEG — electrical brain activity recorded from the scalp — is one of the hardest real-world signal modalities in ML: low signal-to-noise ratio, massive subject variability, and device inconsistencies. Most people avoid it for these reasons. As a research intern, you'd be embedded with the founding team, working on problems like: Running and analyzing experiments on self-supervised pretraining objectives for EEG data across subjects, devices, and recording conditions Benchmarking and stress-testing existing approaches (like recent EEG foundation model papers) to understand exactly where and why they break: cross-dataset, cross-montage, under distribution shift Building and refining evaluation protocols that distinguish real progress from noise, so we're not fooling ourselves with leaky benchmarks Writing internal research memos that become the shared knowledge base of the lab — "why model X fails on dataset Y," "what Z dataset teaches us incorrectly," "what we tried and why it didn't work" You'll gain deep exposure to a modality that very few people have studied with modern ML tools — and your contributions will directly compound into the company's core technology. Where this is going We're building toward a world where thought is an interface. You silently compose a message and it types itself. You navigate an AR display without lifting a finger. Software adapts to your cognitive state in real time. A universal interface between human thought and digital action. The product we're building to get there has three layers: A Neural Encoder: a foundation model that maps raw EEG into robust, reusable embeddings that work across devices, subjects, and contexts A Neural API: a stable interface that any app can call: "What is the user's state?" "What intent is most likely?" "What changed?" Reference applications: proving utility and driving our data collection flywheel Near-term, the use cases are already real. A limited vocabulary of thought-to-action commands (volume, select, activate, navigate) would feel like magic to consumers. Sleep staging, stress detection, cognitive load monitoring, and engagement measurement are all feasible with today's signal quality. On the clinical side, we're pursuing avenues like epilepsy monitoring and migraine pre-emption as a wedge for high-quality data, credibility, and early revenue. Hardware matters too. No comfortable, discreet consumer device today covers the brain regions needed for language decoding. We'll eventually design our own. Think a normal-looking baseball cap with dry electrodes hidden in the brim, or something that looks more like AirPods than a medical device. The model needs to be hardware-agnostic, because the form factors will keep evolving. Research culture We have a few strong opinions about how research should work: Minimal hand-engineering, maximal learning pressure. We're skeptical of approaches that hard-code domain heuristics into the model. We'd rather let models discover structure than force-feed it. If you've read Sutton's Bitter Lesson and felt something click, we're on the same page. Reproducibility over vibes. If we can't answer "which preprocessing version produced this result," we don't trust the result. Every experiment is tracked, every pipeline is versioned, every claim is stress-tested. Internal criticism is encouraged. The fastest way to build real knowledge is to kill bad ideas early. We want people who are comfortable saying "I think this is wrong." Failed experiments are documentation, not waste. We write up what doesn't work with the same care as what does.

Requirements

You're a graduate student (Master's or PhD) — or an exceptional undergraduate — studying machine learning, computer science, statistics, physics, neuroscience, or a related field.
You're excited about EEG and brain-computer interfaces, and you want to do real research, not busywork.
Coursework or research experience in deep learning, with exposure to self-supervised or representation learning
The ability to implement, run, and debug ML experiments independently (PyTorch preferred)
Comfort working with messy, heterogeneous data
Strong written communication — you can clearly explain what you tried, what worked, and what didn't
Familiarity with (or eagerness to learn) the EEG landscape: public datasets, benchmarks, and where current approaches fall short

Nice To Haves

Experience with EEG or neural signal decoding is a strong plus — if you've already worked with this data, you know how messy and rewarding it is.
Experience in EEG/BCI competitions is a great signal.

Responsibilities

Running and analyzing experiments on self-supervised pretraining objectives for EEG data across subjects, devices, and recording conditions
Benchmarking and stress-testing existing approaches (like recent EEG foundation model papers) to understand exactly where and why they break: cross-dataset, cross-montage, under distribution shift
Building and refining evaluation protocols that distinguish real progress from noise, so we're not fooling ourselves with leaky benchmarks
Writing internal research memos that become the shared knowledge base of the lab — "why model X fails on dataset Y," "what Z dataset teaches us incorrectly," "what we tried and why it didn't work"