Senior Machine Learning Engineer

ABOUT ACORAI

Acorai is a Swedish medtech company building the SAVE Sensor System — a non-invasive, handheld device that estimates intracardiac filling pressures at the bedside. It records several sensor channels at once — ECG, PPG, heart sounds and chest-wall motion — and infers, from those signals alone, a pressure that today can only be measured by threading a catheter into the heart.

That inference is a machine learning problem, and it is the core of the company. Heart failure is one of the largest causes of hospitalisation worldwide, and most readmissions are congestion that was not seen in time. If our model works, congestion gets treated before it becomes an admission.

We are validating our algorithm against invasive reference measurements and preparing our US regulatory submission. We are hiring a Senior Machine Learning Engineer to join the team that builds it.

THE ROLE

You will join our existing ML team as a senior individual contributor, working on the model that the product is. This is deep, hands-on work on hard data: short multichannel physiological recordings, expensive invasive labels, small sample sizes, and a bar set by a regulator rather than a leaderboard.

WHAT YOU WILL DO

• Build and improve the models that turn multimodal sensor recordings into estimates of cardiac filling pressure, trained against invasive reference measurements from real patients
• Own your work across the pipeline — signal preprocessing, representation and feature learning, training, thresholding and calibration, signal-quality gating and indeterminate-output handling
• Design experiments that answer questions rather than produce numbers: learning curves, ablations, subgroup analyses, leakage audits, participant- and site-level partitioning
• Characterise the model honestly — performance across demographic and acquisition subgroups, robustness to sensor noise and acquisition variability, calibration, and drift over time
• Write the algorithm and dataset documentation that goes into FDA submissions: model description, training/tuning/validation dataset provenance and representativeness, performance characterisation, and predetermined change control plans
• Take models from research to a locked, versioned, deployable artefact — including inference under embedded hardware constraints
• Work with our clinical and regulatory teams on what data to collect next and what it is actually worth

WHAT YOU BRING

• 5+ years building machine learning systems that shipped, with substantial depth in time-series or signal data
• Physiological or sensor time series — ECG, PPG, accelerometry, IMU, acoustic, or similar. You understand why biological signals are not text or images and why most off-the-shelf recipes underperform on them
• Serious ML practice: you build partitions that do not leak, you know why a model that looks excellent on a random split fails at a new site, you calibrate, you quantify uncertainty, and you are sceptical of your own results before a reviewer is
• Strong Python and PyTorch (or equivalent); reproducible training, experiment tracking, versioned data and models
• Comfort with small-n and expensive labels — we cannot simply collect more
• Startup temperament: you own problems end to end, build the tooling you need, and are comfortable that some of the answer does not exist yet

NICE TO HAVE

• Regulated medical device ML (SaMD) — algorithm documentation for a 510(k), De Novo or PMA, PCCP, IEC 62304, ISO 14971
• Healthcare or clinical data experience, particularly cardiovascular
• Self-supervised or representation learning on large unlabelled signal archives
• Embedded or edge inference — quantisation, latency and memory constraints
• Digital signal processing depth: filtering, denoising, segmentation, beat detection, multi-channel synchronisation
• Published work in physiological signal processing or clinical ML

WHAT WE OFFER

• A model that is the product, not a feature of it
• Real clinical data with invasive ground truth — rare, expensive, and the reason this problem is tractable at all
• Direct influence on what data we collect next and how the evidence is built
• Hybrid working in Portugal, in a team spread across Sweden, Portugal and the US
• Competitive salary and participation in our employee option programme