Radiology Annotation Jobs (NOW HIRING)

We build clinical AI that reads alongside radiologists. Our abdomen-pelvis CT triage device is FDA-cleared, and it's the first commercial system to simultaneously triage seven urgent conditions on abdomen-pelvis CT in the U.S. We’re backed by Khosla Ventures.

This role owns two of the things that decide how good our models can get: the quality of the labels going in, and whether a radiologist can see and trust what the model gives back.

Radiologist time is the most expensive input we have. When a reader has to click four times to do something that should take one, we lose annotation throughput, and less throughput means weaker models, which eventually means a finding a patient's scan should have caught. So the interface a radiologist works in genuinely drives model quality, and this is a product engineering job as much as an infrastructure one.

The harder half is knowing whether the labels are any good in the first place. Our annotation pipeline is built to measure itself: cases are claimed without race conditions, annotators move through defined phases, some batches are seeded with known ground truth, others are handed to more than one reader on purpose, and we score agreement with per-lesion Dice even when two readers worked from reconstructions that don't share a geometry. Getting that measurement right is most of the work.

Today one engineer holds this whole surface while also carrying several others, and that's the gap we're hiring to close.

The viewer, built on Cornerstone3D and VTK.js. A unified volume-rendering path that falls back to stack rendering, multiplanar reformats generated on demand in any plane and clearly labeled as reformats, and progressive loading that builds a low-resolution volume from the first 10% or so of each HTJ2K codestream so the reader sees an image right away. It also includes the correctness work that only shows up when it's wrong: radiological left/right, rulers under gantry tilt, MONOCHROME1 inversion, and signed-pixel codec mismatches.

The annotation system. 3D mask storage, AI-assisted click-to-segment across all three planes, the classical tools readers still reach for (region growing, FWHM thresholding, multi-seed refinement), and the evaluation pipeline around them: case-pool ledgers, per-annotator phase state machines, ground-truth and peer-overlap batches, agreement scorecards, and cross-series resampling through NIfTI affines. One thing we're firm on: when a case can't be scored, the system says so, rather than recording it as a zero.

Annotation schemas. Per-lesion-category schemas with gating by view, phase, and slice, and conditionally required fields. These calls are partly clinical, and you'll make them together with our radiologists.

Model output that a radiologist can actually read. A versioned sparse-RLE mask contract keyed by SOP Instance UID, and the geometry that maps a model's 512-square grid through image position, direction cosines, and pixel spacing into world-space contours that land on the right anatomy.

The platform underneath. FastAPI on AWS, with per-study authorization enforced on every read, cohort isolation between customers, and access-trail auditing that meets HIPAA §164.312(b) for study-data reads.

A lot of the hard bugs here come down to DICOM geometry: coordinate systems, orientation, affines. You don't need to arrive knowing all of it, but you should find it interesting rather than tedious, because you'll spend real time in it.

Someone who has built software people used for hours a day and made it better by watching them use it. When you're asked whether the annotation quality is good, your instinct is to reach for a metric. You're comfortable enough with coordinate systems to track down why a mask is 3mm off instead of nudging it into place, and you fail safely by default when patient data is involved. You can disagree with a radiologist about software, and you defer to them completely on medicine.

React and TypeScript with real performance work behind you; backend API and async experience; medical imaging (DICOM, Cornerstone3D, OHIF, PACS); annotation tooling from either side; AWS and Terraform; segmentation or computer vision; inter-rater agreement and measurement design; regulated software experience (ISO 13485, IEC 62304, HIPAA); and codec or streaming work such as HTJ2K.

React, TypeScript, Vite, Cornerstone3D, VTK.js, Jotai, TanStack Query, Tailwind, Playwright; Python, FastAPI, pytest; AWS (DynamoDB, S3 byte-range reads, Cognito, SQS, ECS, Lambda) with Terraform; DICOM, DICOMweb, HTJ2K, NIfTI, RLE.

$140,000 to $180,000 base, plus an approximately 10% discretionary bonus. Equity may be offered to top candidates.

for candidates outside the US, the base is cash-weighted and set by the local market for the country where the work is done, and we'll share the specific range for your location early in the process. You would work from your own country, so no visa is needed and there are no immigration strings. International offers are cash-only, with no equity.

several of our core repositories are already owned by engineers outside the US, and you'd have the same repositories, data, and review authority as anyone on the team. Our reviews are written and asynchronous because the time-zone spread calls for it, which happens to suit regulated code well. Depending on your country, you'd join as a contractor or as an employee through an employer of record. We don't run a two-tier engineering team.

a2z was co-founded by Pranav Rajpurkar, an Associate Professor at Harvard Medical School with more than 150 publications. Our engineers trained at MIT and Stanford, and our fellowship-trained radiologists read alongside the model every day.