Embedded Machine Learning Engineer

Embedded Machine Learning and Real-Time Sensor Classification

Client

Location:

Edinburgh

Salary: £60,000 to £70,000 per annum

Our client is building the next generation of real-time detection systems that operate at the edge of the network, where connectivity is unreliable and power is constrained. Their vision is straightforward and radical: machine learning models that run on embedded devices, processing sensor data in the field, making instant decisions without relying on cloud infrastructure or continuous data transmission.

This is not a supervised learning problem on tabular business data. This is not model serving from a GPU cluster. This is embedded machine learning in its most demanding form: taking sophisticated sensor-based classification systems and making them run reliably on devices with megabytes of RAM, in real-world environments where data is noisy, conditions are uncontrolled, and failure is not an option.

The client is at the frontier of what embedded ML can do. Most organisations are still building cloud-first systems. They are building ground-truth systems: devices that operate autonomously, make decisions in real time, and remain reliable under the constraints that define actual field deployment.

The engineering challenge is significant. You will be responsible for the complete lifecycle of machine learning models that run on embedded devices. This means:

• From Sensor to Deployed Device. You will work with raw sensor data streams from physical devices operating in uncontrolled environments. This data is not clean. It carries noise, calibration drift, temperature sensitivity, and the unpredictable behaviour of systems deployed in the real world. Your job is to transform this stream into a robust, real-time classification model that runs on hardware with severe constraints on memory, computation, and power.
• Engineering Under Constraint. Every machine learning decision you make has downstream consequences for embedded systems. A model that requires 512MB of RAM will not run on a device with 128MB. A model that takes 500ms per inference will drain the battery in days. A model that works in the lab but degrades in the field is a failure. You will learn to think like an embedded systems engineer, not a researcher.
  
  You will understand that the goal is not maximum accuracy; it is maximum accuracy within the hardware constraints that define your actual platform.
• Real-World Iteration. You will take models from lab development through semi-field validation and into live deployment. You will see what happens when your model encounters conditions it has never seen before. You will debug why a model that performed flawlessly during development is behaving unexpectedly in the field. You will iterate based on real-world performance data and build the judgment to know when a model is sufficiently reliable for deployment.
• Cross-Domain Transfer. The long-term vision is building a platform where machine learning models and data structures developed for one application can be adapted and transferred to others. You will contribute to demonstrating that this is technically possible; that you can take a model built for one classification task and, with appropriate retraining and adaptation, make it work reliably in a different domain.
  
  This is the frontier of applied ML.

Reporting to the Lead Data Scientist, you will own the development and deployment of machine learning models that sit at the heart of the embedded detection platform. You will work closely with R&D scientists, firmware engineers, hardware designers, and product teams. You will be the bridge between data science and embedded systems. You will own the complete pipeline: from sensor data ingestion and cleaning, through feature engineering and model development, through embedded optimisation and deployment, through field validation and iteration.

• Clean, structure, and analyse sensor datasets from real-world deployments for training and evaluation
• Develop machine learning models optimised for embedded deployment on resource-constrained devices
• Work with Tensor Flow Lite, Edge Impulse, or custom firmware deployment strategies to integrate models into actual hardware platforms
• Optimise model architecture, quantization, and inference speed to fit hardware constraints without unacceptable loss of accuracy
• Collaborate with firmware engineers to integrate your models into live devices, understanding and accommodating their constraints
• Test model performance across lab, semi-field, and real-world settings; identify failure modes and iterate
• Document training pipelines, feature engineering methods, model validation results, and deployment learnings
• Establish reproducible, scalable workflows for model development, retraining, and versioning
• Support future platform expansion by demonstrating that models can be adapted across different applications and use cases
• Develop model-driven features such as confidence scoring,…