Head of Machine Learning

Empathy Talent; San Francisco (CA); On-site; USD 120,000 - 200,000; Full time

Job summary

Empathy Talent in San Francisco is seeking a Machine Learning Engineer to own a challenging project: turning noisy physiological signals into a robust, continuous language interface. You will lead data collection, signal representation, and evaluation with founders and hardware leads.

You will develop generalizable representations from unlabeled data and build sequence models that generalize across users and devices, with ~15 minutes of calibration data.

Qualifications

• Deep experience applying ML to noisy, distribution-shift problems.
• Backgrounds may include: ASR / Speech ML, Audio and speech modeling, Time-series modeling, Brain-Computer Interfaces (BCI), Radar or sensor-based ML, Multimodal or physiological signal processing.
• PhD-level research ability preferred.
• Production-quality code, ability to design experiments independently, and adapt quickly to changing directions.

Responsibilities

• Lead end-to-end ML efforts, collaborating with founders and sensing and hardware leads to define data collection, signal representation, model families, and evaluation.
• Learn generalizable representations from large amounts of unlabeled and weakly labeled physiological time-series data.
• Build continuous sequence models using approaches such as Conformers, Mamba-style architectures, CTC, transducers, and pretrained speech/language models.
• Adapt large cross-user models to new users using approximately 15 minutes of calibration data.
• Design rigorous user-held-out, session-held-out, and device-held-out evaluation frameworks.
• Scale model training across thousands of hours of data and thousands of users.
• Optimize the complete ML system for continuous, low-latency, real-time operation.
• Help define the broader ML architecture, research direction, and technical roadmap as the team scales.

Job description

We are looking for a Machine Learning Engineer to own one of the company’s most important technical challenges: turning weak, noisy, and highly variable physiological signals into continuous language.

The team has already built prototypes capable of decoding subvocal speech at more 200 words per minute. The next—and significantly harder—challenge is generalization.

A model that performs well for one person, during one session, with one device placement is not enough. The system needs to remain accurate when that person returns the next day, when hardware placement changes slightly, and eventually when an entirely new user puts on the device with only a few minutes of calibration data.

You will lead this effort end to end, working directly with the founders and sensing and hardware leads to determine what data should be collected, how signals should be represented, which model families should be pursued, and how progress should be evaluated.

What You’ll Work On

• Learn generalizable representations from large amounts of unlabeled and weakly labeled physiological time-series data.
• Build continuous sequence models using approaches such as Conformers, Mamba-style architectures, CTC, transducers, and pretrained speech/language models.
• Separate speech-related information from variations caused by anatomy, device placement, session, hardware, and environmental conditions.
• Adapt large cross-user models to new users using approximately 15 minutes of calibration data.
• Design rigorous user-held-out, session-held-out, and device-held-out evaluation frameworks.
• Scale model training across thousands of hours of data and thousands of users.
• Optimize the complete ML system for continuous, low-latency, real-time operation.
• Help define the broader ML architecture, research direction, and technical roadmap as the team scales.

Technical Environment

The current ML stack is primarily:

The team also maintains custom infrastructure supporting signal processing, data collection, experiment tracking, training, and evaluation.

What We’re Looking For

You may be a strong fit if you have deep experience applying machine learning to challenging signal-based problems where the data is noisy and distributions shift frequently.

Relevant backgrounds may include:

• Automatic Speech Recognition (ASR) / Speech ML
• Audio and speech modeling
• Time-series modeling
• Brain-Computer Interfaces (BCI)
• Radar or sensor-based ML
• Multimodal or physiological signal processing
• Other signal domains involving significant variability and distribution shift

We are particularly interested in candidates with PhD-level research ability, whether or not that experience came through a formal PhD program.

Strong candidates should also be comfortable writing production-quality code, designing experiments independently, and moving quickly as research findings change the technical direction.

Why This Role Is Different

This is not a position where you will inherit an established model architecture and spend your time making incremental improvements.

You will help determine how the ML problem itself should be framed.

You will have significant ownership over the modeling strategy, data strategy, evaluation methodology, and technical direction while helping build the initial ML organization around you.

The work you do will directly influence whether an ambitious research breakthrough can become a reliable, real-world product.

Location: Hybrid in San Francisco, CA

Base Compensation: Up to $200K, with some flexibility

Equity: 0.50%–2.00%

Work Authorization: Must be a U.S. Citizen or Green Card holder