AI Engineer for Advanced Sports Analytics Project

As a Machine Learning Engineer, you will embark on a challenging yet rewarding project aimed at developing algorithms to recognize exercise movements using Inertial Measurement Unit (IMU) data. This role requires a deep understanding of data engineering, AI, and the nuances of human motion. You will design sophisticated models to identify and classify hundreds of exercise movements with a focus on achieving a 99.9% accuracy rate. Additionally, you will refine these models to operate efficiently on consumer devices like mobile phones and watches, maximizing the use of device-specific features such as GPU and NPU.

Моя компания специализируется на разработке мобильных приложений и ML решений на заказ, преимущественно для рынка США. У ребят из команды есть опыт работы в Google, Сбере и других больших технологических компаниях, а у компании разработки приложений как для успешных стартапов, так и для устоявшихся компаний. У нас положительный опыт найма и сотрудничества с коллегами из чата. Сейчас у нашего партнера, HiQ Fitness, для которого мы активно ведем разработку, открыты две вакансии: AI Engineer и LLM Engineer в проекте по спортивной аналитике. Есть возможность работы как через мою компанию так и напрямую.

HiQ Fitness is on a mission to utilize truly sophisticated and rigorous artificial intelligence to accelerate the pace of improvement in athletic performance. In short, athletes who use our technology will ‘Get Better Faster’ and unlock a powerful competitive edge that has never been available.

• Clean and preprocess IMU data, which may include transforming data to a common reference frame using quaternions.
• Develop intelligent methods to recalibrate sensors as needed without user intervention.
• Develop advanced feature extraction techniques for time-series data.
• Experiment with a variety of temporal machine learning models, emphasizing those (like MobileNet and U-Net) that leverage native GPU and NPU abilities present on today’s mobile devices to determine the best approach for classification of our data.
• It is possible that LSTMs, CNNs, Time-Distributed CNNs + RNNs, Transformers, Hidden Markov Models (HMMs), or ensemble models could be helpful.
• Architect and improve secondary models that transform athlete state recognition into ‘rep counts’ and ultimately govern the behavior of the model.
• Build recognition model & workout execution model validation tools to measure performance based on accuracy, precision, recall, and F1 of the model’s overall and micro performance.
• Analyze deficiencies and collaborate to improve performance using increasingly rigorous approaches in the short, medium, and long term.
• Develop an effective model evaluation scheme and method for determining which sensors are needed for accurate movement recognition. Our goal is 99.9%+ accuracy in rep counting.
• We believe that the key to delivering this level of accuracy will be accurately determining which movements will need to be disambiguated by users vs. those that can be detected without user intervention. (Example: Wall Balls and Thrusters look very similar when looking at data from IMUs placed on foot, waist, and wrist.)
• Architect and implement dev-to-production deployment strategy.
• Models to be deployed to Android and iOS phone and smart watch devices.
• Other models may be deployed to the proprietary firmware stack (Nordic nrf52840 MCU residing on Seed XIAO nrf52840 Sense boards)
• Document the research findings & development process clearly and comprehensively.
• Collaborate with our team to understand project requirements and contribute technical insights.

• Degree (ideally Master’s or more) in Computer Science, Data Science, Machine Learning, or a related field.
• Expert Proficiency in Python and deep experience with machine learning libraries such as TensorFlow ( and TensorFlow Lite), PyTorch, and Keras.
• Prior work with RNN family of temporal learning models as well as the other modeling paradigms discussed in the responsibilities section above. In general, material exposure to numerical sequence based, preferably in a similar context.
• Understanding of sequence models and their implementations.
• Knowledge of data preprocessing, feature extraction, and model evaluation techniques for time-series data.
• Familiarity with handling IMU data and quaternion transformation for data normalization and augmentation.
• Experienced, capable, & successful; but shows up each day as genuinely humble, collaborative, and eager to learn.
• Strong problem-solving skills.
• Completely fluent and comfortable working in English
• Strong communication and documentation skills, capable of working collaboratively in a multidisciplinary team to translate complex product requirements into actionable development plans.
• A deep understanding and appreciation for CrossFit workout paradigms and movements.
• The ideal candidate will be an active CrossFit Athlete who takes classes in a CrossFit box at least 3 times per week.
• The ideal candidate will also have a broad, applied Data Science skillset that can be used (either directly or as a mentor or, if qualified, a manager to others on the team) to tackle the many other data science problems we plan to tackle outside of the ‘movement recognition’ project.
• Standard predictive modeling techniques (Regression, etc.)
• LLM fine-tuning, RAG, and LangChain style LLM workflows that trigger application logic (i.e. GPT Functions).

Бонусы длительного сотрудничества с моей компанией:

• Помощь в создании и продвижении профиля на Upwork.
• Рост часовой ставки со стажем.
• Возможность перехода на full-time позицию.