SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Throne is a continuous tracking device for personalized insight into gut health and hydration, co-founded by John Capodilupo (co-founder and former CTO of WHOOP). The company applies continuous health tracking rigor to signals that have traditionally been guesswork, with a north star of improving health and saving lives.
You will own highly challenging sensing and machine-learning problems end-to-end, from definition through deployment and continuous improvement. This is not a narrowly scoped role—you'll be expected to understand product objectives, oversee the data ecosystem, identify highest-leverage technical work, make sound tradeoffs, and drive solutions through implementation, validation, and deployment.
Key Responsibilities:
- End-to-end solution ownership: take loosely defined product or sensing problems, determine what needs to be learned or built, identify critical unknowns, develop a plan, execute it, and drive work to measurable outcomes.
- Own model performance across computer vision, video, audio, and next-generation sensing systems, including classification, segmentation, detection, embeddings, and temporal/video understanding.
- Large-scale dataset development: data mining, labeling strategy, dataset quality, split design, active learning, failure analysis, and evaluation construction.
- Model training and pre-training, including fine-tuning, transfer learning, self-supervised and semi-supervised approaches. Pre-training base CV models experience is a strong plus.
- Model efficiency work: quantization, distillation, pruning, architecture selection, ONNX/TorchScript tradeoffs, latency and memory reduction, cost/performance optimization.
- Experimentation and evaluation rigor: metrics, ablations, cohort analysis, error taxonomies, regression tests, confidence calibration, and clear go/no-go criteria.
- Sensor validation for next-generation devices using scientific approaches to compare sensors, characterize signal quality, design experiments, and determine whether new modalities improve real-world model performance.
- Data science and SQL workflows to inspect production data, evaluate model behavior, build analysis datasets, and connect model outcomes to product and device behavior.
- Production ML integration with backend, firmware, and data systems to ensure models are deployable, observable, reproducible, and reliable.
- Research translation: stay current with relevant CV, video, multimodal, compression, and foundation-model research, turning useful ideas into measurable improvements.
What Makes This Role Different:
Throne is not looking for someone waiting for a precisely scoped modeling problem. You'll architect and implement programs from question to concept to production. You might start with "Why is this model underperforming?" or "Will this new sensor materially improve our ability to predict a clinical outcome?" From there, you'll own data collection, curation and analysis, failure mode investigation, labeling and third-party labeler management, experiment design, model updates, inference pipeline modifications, production instrument behavior understanding, and challenging legacy assumptions—all with cost-effective, production-quality product in mind. You will own Throne's AI/ML ecosystem.
Requirements:
- 7+ years of applied ML experience with strong production or research-to-production ownership.
- Deep experience training and improving computer vision models across multiple domains (not just using off-the-shelf APIs).
- Hands-on experience with pre-training or adapting base CV models, including dataset design, scaling tradeoffs, and representation quality.
- Strong PyTorch experience and comfort debugging the full model stack: data loading, augmentation, loss design, training stability, metrics, deployment export, and inference performance.
- Demonstrated experience with model optimization: quantization, distillation, efficient architectures, latency/cost tuning, and accuracy-efficiency tradeoffs.
- Strong SQL, Python, and data science skills for working directly with large datasets, labels, metrics, and model failure cases.
- Ability to work across ambiguous sensing problems with scientific discipline: form hypotheses, design experiments, control variables, and make evidence-backed recommendations.
- Capable of working well with modern LLM tools to scale research, coding, analysis, and experimentation efficiently.
- Self-starter mindset: able to improve the stack from data, labels, modeling, evaluation, deployment, and instrumentation rather than waiting for narrowly scoped training tasks.
- Strong communication and technical judgment, especially around deciding what model work is likely to move real product performance.
Nice to Have:
- Experience with video models, segmentation models, embeddings, temporal classification, multimodal learning, audio/signal processing, or sensor-fusion systems.
- Experience with labeling tools (Supervisely, Roboflow, CVAT, etc.), MLflow, W&B, Optuna, SageMaker, TorchServe, ONNX Runtime, TensorRT, or similar ML infrastructure.
- Experience deploying models under edge, embedded, mobile, GPU, or strict latency/cost constraints.
- Experience with labeling pipelines, active learning, synthetic data, weak supervision, human review workflows, or model-assisted annotation.
- Experience validating new hardware sensors or working closely with firmware/device teams.
