Member of Technical Staff, Physical AI

The opportunity

Substrate is building a laboratory that runs itself. Robotics move the samples and the instruments take the measurements, but a lab that operates without people standing over it has to perceive what it is doing: every camera frame, every sensor stream, every instrument log, captured and understood in real time. That perception layer does not exist yet. You would build it.

This is the work that lets an autonomous lab catch its own mistakes, a misfired liquid handler or a transfer that did not complete, the moment they happen, and that captures the full provenance of every experiment as it runs. Provenance is one half of the bar, with scientific quality, that makes Substrate's data worth training on, and it only exists if it is captured at the source. The layer has to be live before the lab starts running semi-automated, because data that is not captured cannot be recovered.

About Substrate

Substrate is building the critical infrastructure layer between AI and biology: an AI-native automated lab that produces biological data at scale. AI for biology has a data problem, not a compute problem. Biological foundation models can predict, but they cannot run experiments, and the high-quality, large-scale data they need does not exist. Substrate generates it, with quality and provenance built in.

We are four co-founders, venture-backed, building our first lab at 20 Triton Street in London. We are not a cloud lab and we are not a CRO. We are the infrastructure that turns scientific intent into executed experiments and structured, AI-ready data, and over time into proprietary datasets and our own infrastructural intelligence.

The role

You will own how the lab perceives itself, end to end. That means onboarding the sensors and cameras across the lab and turning their output into a single, timestamped data model, mapped to a live picture of the physical space. It means standing up vision models (VLMs and VLAs) that watch the lab run and detect errors in real time, so the team is alerted the instant something deviates from what the workflow expected. And it means doing this on edge compute on-site, because real-time error detection cannot wait on a round trip to the cloud.

Your capture layer is the foundation the rest of the software builds on. The infrastructure software turns customer intent into executed lab work; the intelligence layer learns from what the lab produces. Both depend on clean, complete, well-structured data coming off the floor, which is what you build. You will work day to day with the founding software engineer who owns the infrastructure software, with the scientists running the assays, and at the boundary with the intelligence team who take your data into the metadata layer.

Your first 90 days

FIRST 30 DAYS

◆ Get the first cameras and sensors capturing on the live, semi-automated lab, so no run happens uncaptured.

◆ Agree the timestamped, mapped data model with the software and science teams.

DAYS 30 TO 60

◆ Ship the first real-time error detection on a live instrument, alerting the team the moment a run deviates.

◆ Extend capture to instrument telemetry and the movement of materials across the lab.

DAYS 60 TO 90

◆ Add environmental sensors (thermal, humidity) and harden the edge pipeline and the detection models.

◆ Feed clean, mapped data into the metadata layer the intelligence team builds on, ready to scale toward full automation.

Who you are

You have built perception or sensing systems that had to work in the physical world, where the interesting failure modes live: robotics, autonomous vehicles, drones, manufacturing, or scientific instruments. You have made cameras and sensors behave (calibration, synchronisation, capture, deployment) and you have put computer vision into production, not just prototypes. You are pragmatic about vision-language models: you have used them in real systems and you know where they break.

You are happy being the person who makes the physical thing actually work, hands on, as an individual contributor. You are comfortable with early-stage ambiguity, with deciding on partial information, and with revisiting decisions when better information arrives. You want to be close to the hardware and close to the science.

MUST HAVE

◆ Experience building real-time perception or sensing systems in the physical world (robotics, autonomous vehicles, drones, manufacturing, or instrumentation).

◆ Hands-on experience with cameras and sensors: calibration, synchronisation, capture pipelines, and deployment.

◆ Production computer vision experience, including vision-language models (VLMs or VLAs) or strong classical CV.

◆ Strong software engineering, and real comfort at the software-to-hardware boundary.

◆ Four or more years building systems of this kind.

NICE TO HAVE

◆ Edge deployment and real-time inference on on-site compute (for example Nvidia Jetson or similar).

◆ Sensor fusion, SLAM, or large-scale time-series and telemetry.

◆ Exposure to lab automation, scientific instruments, or other regulated physical-data environments.

◆ Early-stage or founding-engineer experience at a venture-backed company.

Why this is unusual

Most physical-AI roles sit inside robotics or autonomous-vehicle companies, where perception is the product. This is different. Here, perception is the sense organ of a laboratory that runs itself. You will not be detecting pedestrians; you will be catching a pipetting error before it corrupts a dataset that a foundation model will train on, and capturing the provenance that makes that dataset worth training on at all.

You will also sit between software, hardware and biology at the same table. Some engineers find that mix energising; some find it distracting. It is worth knowing in advance which one you are.

How we work

Because the work is hands-on, you will spend significant time in the lab at 20 Triton Street, at a minimum of three days a week, where the instruments, robots and sensors are. The rest of the team is distributed across several locations and works flexibly, and we keep a light shared rhythm: a Monday kickoff, a Thursday all-hands, a short daily team sync, and a quarterly offsite. We are a small team that documents in the open and backs the best idea regardless of who has it.

We look after people well. In the UK that means 30 days of annual leave a year plus public holidays, a pension with a 10% employer contribution, and top-tier private health cover with Bupa, with more added as the team grows.

The team you will join

You will join a founding team of four. Mostafa ElSayed is our CEO, and also the founder and CEO of Automata, the lab-automation company whose hardware our lab runs on. Oli Hoy leads the lab build, the infrastructure software and operations. Alexey Morgunov leads the intelligence software and the science. Anna Huyghues-Despointes leads commercial and partnerships. You will report to our founding software engineer, who owns the infrastructure software, and work alongside the scientists and the intelligence team. We expect to be over 30 people within the year, so you will be early, and it shows in the scope you own.

The process

Screening, then a behavioural and cultural-fit conversation, then a technical session or work sample with the team (including time in person at the London lab), then references.

Substrate is an equal opportunity employer. We make hiring decisions on merit, scope-fit, and the strength of the working relationship we expect to build with each hire. Applications welcome from candidates of any background.