Founding Machine Learning Engineer, Industrial AI in Oxford

This role is being recruited by Calyptus on behalf of one of our clients.

Calyptus is a specialist recruitment partner helping startups and high-growth technology companies hire across engineering, product and design.

• What to expect
• We’ll review your application and, if your experience is a strong match, one of our team will be in touch
• We’ll support you throughout the interview process and keep you updated at each stage
• If this particular role isn’t the right fit, we may consider you for other relevant opportunities with our clients

The role

The core of this job is taking raw sensor data from extrusion lines and building the algorithm that decides what the line should do differently.

It has to be right often enough, and explain itself well enough, that an experienced operator will act on it.

In time it will drive setpoints directly, so it also has to know when it does not know.

Getting the recommendation right is only half the job.

The other half is understanding how and when a recommendation can safely become an actual setpoint change on the machine, and building the supervisory control layer that makes that happen without fighting the line's existing controls or its operators.

You will also be responsible for version one of the product around that algorithm: the data platform, the models, the deployment and the application operators use.

This is a senior, hands‑on role with ownership of the technical stack.

You will make the architectural calls, write the code, spend time on customer sites and help hire the engineers who follow.

What you will do

• The decision-making algorithm
• Design and build the core algorithm that turns live sensor data into decisions: which setpoints to change, by how much and when.
• Frame the problem properly.

Decide whether it is best treated as prediction plus optimisation, model predictive control, Bayesian optimisation, contextual bandits, offline reinforcement learning or a combination, and justify the choice with evidence.

• Build process models that link settings and operating conditions to quality, scrap, throughput and energy, and that respect the dynamics of the process, including time delays, dead time and variation in raw material.
• Learn from historical data in which settings were rarely varied deliberately.

Deal with confounding, operator interventions and changes of regime, and design safe experiments on live lines to fill the gaps.

• Quantify uncertainty and build in the ability to make no recommendation when conditions fall outside what the model has seen.
• Evaluate decisions properly: offline policy evaluation before anything goes live, and controlled trials against real production outcomes afterwards.
• Sensor data and features
• Build the pipeline from raw sensor data to model-ready features, covering synchronisation, resampling, lag alignment, filtering, missing data and faulty sensors.
• Develop soft sensors for quality measures that are only taken occasionally or offline.
• Link process data to production context such as product, recipe, material batch and shift.
• Detect anomalies, sensor faults and model drift, and retrain as lines, materials and products change.
• From recommendation to setpoint change
• Work out, for each adjustable variable, when a recommendation can safely be applied, taking account of process state, operating envelope, rate of change, interaction with other variables and the time the process needs to settle.
• Design the supervisory layer that sits above the line's existing PLC control loops, adjusting their setpoints while those loops continue to regulate the process, and make sure it never works against them.
• Define the conditions under which the system must hold or hand control back to the operator, such as start‑up, product changeovers, alarms, sensor faults or unfamiliar material, and make handover between manual and automatic operation smooth.
• Monitor the effect of every change and reverse it automatically if the process does not respond as expected.
• Design the staged route from advisory to closed-loop operation, with clear, measurable criteria for when a line, or an individual variable, is ready for the next stage.
• Specify how setpoints are written back to the PLC (range checks, rate limits, reversibility and fallback to the last known good state), working with our integration partner and the customer's engineers.

Machine safety functions stay in the PLC's safety layer and are never overridden by our software.

• Present recommendations with the reasoning behind them, in a form operators can check quickly and trust.
• Engineering and product
• Set the technical architecture for version one and make the build‑or‑buy decisions.
• Put proper machine learning operations in place: versioned data and models, testing, monitoring, deployment on edge hardware and in the cloud, and rollback.
• Spend time on customer sites to understand the process and earn the confidence of plant engineers.
• Help recruit and lead the engineers who join after you.
• What you will bring
• At least six years in applied machine learning or data science, including substantial work on sensor or time‑series data from physical systems such as manufacturing, energy, process industries, robotics or automotive.
• At least one model you took into production where its outputs drove real operational decisions.
• Strong time‑series modelling, statistics and probabilistic methods, including uncertainty quantification.
• Experience with optimisation or sequential decision‑making, such as model predictive control, Bayesian optimisation, bandits, reinforcement learning or constrained optimisation, and the judgement to know when something simpler will do.
• A good grasp of causal reasoning and of the pitfalls of learning decisions from observational data.
• Production‑quality Python and the standard machine learning stack (Num Py, pandas or Polars, scikit‑learn, Py Torch), with the engineering discipline to match: testing, CI, Docker and Linux.
• A solid understanding of supervisory control: process dynamics, feedback, PID, dead time and loop interaction, and how a layer that adjusts setpoints interacts with the regulatory loops beneath it.

You should be able to explain why a correct recommendation can still be unsafe to apply.

• Experience of owning a technical product from start to finish, and the ability to explain technical decisions clearly to plant managers, operators and investors.
• Experience on continuous processes such as extrusion, compounding, film, paper, metals or chemicals.
• Offline reinforcement learning, model predictive control or advanced process control running in production.
• Commissioning supervisory or closed‑loop control on a live plant.
• Hybrid or physics‑informed modelling that combines first principles with data.
• Machine learning operations tooling (MLflow or similar), edge inference (ONNX Runtime or similar), time‑series databases and a major cloud platform.
• Working knowledge of PLCs, OPC UA and MQTT, and familiarity with ISO 13849‑1, IEC 62061 and IEC 62443.
• Time at an early‑stage company, or as a founder.
• A degree or Ph D in machine learning, statistics, applied mathematics, control, chemical engineering, physics or a related discipline. Equivalent experience counts just as much.
• Genuine opportunities

Every role advertised by Calyptus is a live position we’re actively recruiting for.

Some clients prefer to remain confidential during the early stages of the hiring process.

When that’s the case, certain company details may be shared later in the process.

Why work with Calyptus?

• Access opportunities with ambitious startups and high‑growth technology companies
• Direct support from experienced specialist recruiters
• Honest feedback and clear communication throughout the hiring process
• Long‑term career support, not just a single job application
• #J-18808-Ljbffr