Neurosoft Bioelectronics is developing next-generation AI for decoding neural time series (high-density subdural ECoG LFPs). We seek an intern machine learning scientist with interest in sequence modelling, state-space methods, self-supervised learning, and/or physics-informed machine learning to build foundation models that infer dexterous finger movements and continuous high DoF upper extremity kinematics, initially from only few minutes of brain data. The role spans research into modern system identification, representation learning, and real-time deployment of low-latency inference pipelines.
Tasks
- Research design and benchmarking Kalman filters, RNNs, LSTMs, Structured State Space Models (S4/S5), Mamba, Neural ODEs, and related sequence architectures.
- Research developing large-scale self-supervised pretraining pipelines and subject-specific fine-tuning workflows.
- Research reproducing and exteding published SoTA neural decoding methods, beginning with LSTM baselines and modern SSM architectures.
- Research decoding individual fingers, grasp synergies, continuous gestures, and full 6 DoF hand pose.
- Research deploying streaming inference using ONNX, TensorRT, Triton, and GPU acceleration.
- Research maintainance of reproducible benchmarking, MLOps, and production research infrastructure.
Requirements
- Familiarity with Python, C/C++, Java, MATLAB, LaTeX, HTML, JavaScript, PyTorch, TensorFlow,
- Familiarity with Signal Processing, Machine Learning,
- Familiarity with Animal Care, Rodent Injections, Microscopy, Sterilization Techniques
- Some capabilities in independent research, scientific writing, strategic planning, time management, presentation
- Optional: Theoretical foundations of time-series modelling, system identification, dynamical systems, differential equations, control theory, and Bayesian estimation.
- Optional: Recent implementation of Kalman filters, HMMs, linear/nonlinear SSMs, RNNs, LSTMs, S4/S5, Mamba, Neural ODEs, or Transformer-based sequence models.
- Optional: Any experience with contrastive learning, VAEs, foundation-model pretraining, transfer learning, domain adaptation, and few-shot learning.
- Optional: Proficiency in Python, PyTorch, CUDA, Docker, ONNX, TensorRT, Triton, GitHub, Linux, and MLflow.
Benefits
- Growth: Opportunity to join an ambitious neurotechnology startup with exposure to multiple areas of company operations.
- Culture: A collaborative and dynamic team environment with direct interaction with founders and leadership.
- Flexibility: Hybrid work model with the possibility remotely and at the office in Geneva.
- Compensation: Entry-level salary range of 45'000 - 55’000 CHF per year. First promotional review after 3 months.
- Ownership: Equity participation through the employee option plan may be available.
We view neural decoding as a problem of system identification: learning latent state-space representations that increasingly approximate the underlying continuous dynamical system generating voluntary movement of a particular individual. Your role is about implementing, testing, and benchmarking the methods developed in collaboration with our academic partners.
Due to the novelty of this effort, this is a mandatory intake role across all seniority levels. At any point you take over an essential practice, tech stack, or reach an essential milestone, your compensation and authority will reflect that. Your role scope and growth are entirely metrics-driven and evaluated quarterly. This is a CSO-track role.