ML Research Engineer / Scientist (Boston, hybrid)

A2z Radiology; Boston, Northern (MA, KY); On-site; USD 160,000 - 200,000; Full time

Job summary

A2z Radiology in Boston is seeking researchers to build models that read whole CT studies like radiologists, not just single slices. You will develop foundation models and vision-language systems that map what the model sees to the language radiologists write.

You will train and evaluate at scale on real-world CT data, with models designed to reach FDA submissions, hospital deployments, and patients. You own your experiments end-to-end from idea to operating point that ships.

Qualifications

• Foundation models that read an entire CT study, not one slice at a time.
• Vision-language learning that ties what the model sees to the language radiologists actually write.
• One model that catches many urgent findings at once, calibrated to operate safely in front of a radiologist.
• Training and evaluation at a scale most researchers never touch.
• Research that ships: your models reach FDA submissions, hospital deployments, and patients.
• You own your models and experiments end to end, from first idea to the operating point that ships.

Responsibilities

• Build foundation models that read an entire CT study, not one slice at a time.
• Develop vision-language alignment between visuals and radiology reports.
• Create a unified model capturing multiple urgent findings with safety in mind.
• Train and evaluate at scale on real-world CT data.
• Ship models that reach FDA submissions, deployments, and patients.

Job description

Build the model that reads the whole scan.

For over a decade, radiology AI has meant narrow tools that flag one finding and stop. We think that is a dead end. We are building the model that reads a CT the way a radiologist does: the whole study, drafted into the report. Your job is to make that model better, faster than anyone expects.

This is research with a short line to patients. The models you train go through FDA clearance and deploy into live hospital reading rooms. The experiment you run this month can change how real studies get read next quarter.

What you\'ll work on

• Foundation models that read an entire CT study, not one slice at a time

• Vision-language learning that ties what the model sees to the language radiologists actually write

• One model that catches many urgent findings at once, calibrated to operate safely in front of a radiologist

• Training and evaluation at a scale most researchers never get to touch

• Research that ships: your models reach FDA submissions, hospital deployments, and patients

You own your models and experiments end to end, from first idea to the operating point that ships.

The data

One of the most exciting datasets in medical AI: real-world CT at ten-million-patient scale, paired with radiology reports and read by our own fellowship-trained radiologists. The raw material most labs only wish they had, and you would train on it from day one.

You'll thrive here if

• You reach for PyTorch the way most people reach for a notebook: custom architectures, training loops, distributed training

• You understand why a loss or an architecture works, not just how to call it

• You design and run your own experiments without waiting to be told

• You care about rigor and reproducibility, because these models carry clinical weight

Helpful, not required

• Medical imaging, 3D, or volumetric data

• Vision-language models or self-supervised learning

• DICOM, CT, or radiology background

• A PhD or publications (a real plus, never a gate)

You do not need prior medical-AI experience. You will learn radiology from the clinicians who read alongside the model every day.

Why now

We have FDA-cleared AI: the first commercial system to simultaneously triage seven urgent conditions on abdomen-pelvis CT in the US. From here the path runs to deployment nationally, and soon internationally. We are backed by Khosla Ventures, and we are a small team that ships. Every model you train stays close to the metal: it reaches submissions, deployments, and patients within months. This is the moment clinical AI grows up, and you would have a front-row seat with your hands on the model.

The team

This is a team of builders and researchers. The same people who created CheXpert, CheXZero, and other widely used open datasets and algorithms in radiology AI have shipped FDA-cleared products into live clinical use. Day to day you would work with ML and software engineers and fellowship-trained radiologists across chest, body, MSK, neuro, and oncology who read alongside the model. a2z was co-founded by Pranav Rajpurkar, one of the most-cited researchers in medical AI.

Details

• Location: Boston (hybrid)

• Compensation: $160,000–$200,000 base + ~10% bonus. Equity may be offered to top candidates.

• Work authorization: Must be authorized to work in the United States

One question we would love your take on when you apply: tell us about a model you built, and what you would change if you trained it again.