Clinical Imaging Platform Engineer

a2z Radiology AI · Boston (hybrid) or Remote (international)

We're building 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. Backed by Khosla Ventures.

Every model we train is limited by two things this role owns: the labels going in, and whether a radiologist can see and trust what comes out.

Radiologist time is the most expensive input we have. A reader who clicks four times for something that should take one click costs us annotation throughput, which costs us model quality, which costs a patient a finding. That chain is short and real. The interface is the throughput.

The other half is harder to see from outside. Generating labels isn't enough, you have to know if they're any good. Our annotation pipeline measures itself: race-free case claiming, per-annotator phase progression, batches seeded with known ground truth, other batches deliberately overlapped between readers, and agreement scored with per-lesion Dice across series that don't share a reconstruction geometry. That's measurement design, not CRUD.

What you'd own:

• The viewer. Cornerstone3D and VTK.js. Unified volume rendering with stack fallback, on-demand reformats in any plane (correctly labeled, never passed off as the source acquisition), and progressive loading that builds a low-res companion volume from the first ~10% of each HTJ2K codestream so a reader sees something immediately. Plus the correctness work nobody notices until it's wrong: radiological L/R, rulers under gantry tilt, MONOCHROME1 inversion, signed-pixel codec mismatches.

• The annotation factory. 3D mask storage, AI-assisted click-to-segment in all three planes, classical tools that earn their place (region grow, FWHM thresholding, multi-seed refinement), and the evaluated pipeline around them: case-pool ledgers, phase state machines, golden-GT and peer-overlap batches, agreement scorecards, cross-series resampling through NIfTI affines. One rule that matters: never a silent zero. A case that couldn't be scored says so.

• Annotation schemas, with clinical judgment. Per-lesion-category schemas, gating by view and phase and slice, conditionally required fields. Where a software decision becomes a clinical one, made with radiologists rather than for them.

• Model output as a product surface. A versioned sparse-RLE mask contract keyed by SOP Instance UID, and the math that scales a 512-square model grid through image position, direction cosines, and pixel spacing into world-space contours landing on the right anatomy.

• The platform underneath. FastAPI on AWS. Per-study authorization as a real primitive, not an internal-user override. Customer cohort isolation. HIPAA §164.312(b) access trails.

You will need to become fluent in DICOM geometry. Almost every hard bug here is a geometry bug wearing a costume. You don't need to arrive knowing it. You need to find it interesting rather than tedious.

Who thrives here:

• You've built something people used for hours a day, and you changed it after watching them. Not shipped and moved on. Watched, then fixed.

• Asked "is the annotation quality good?", your instinct is to define the metric, not describe the process.

• Coordinate systems don't scare you. You'd rather find why the mask is 3mm off than add an offset that makes it look right.

• You default to fail-closed with patient data.

• Enough taste to disagree with a radiologist about software, and enough sense to defer completely about medicine.

Helpful, not required: strong React/TypeScript with real performance work; backend APIs and async; DICOM, Cornerstone3D, OHIF, or PACS; annotation tooling from either side; AWS and Terraform; segmentation or computer vision; inter-rater agreement and measurement design; regulated software (ISO 13485, IEC 62304, HIPAA); HTJ2K or progressive streaming.

Product instinct, measurement instinct, and comfort with hard geometry. That's the bar. Background is a multiplier.

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

Compensation (US, Boston hybrid): $140,000 to $180,000 base plus approximately 10% discretionary bonus. Equity may be offered to top candidates.

Compensation (international, remote): cash-weighted base, tiered by the country the work is performed in. Tier A $110,000 to $150,000 · Tier B $80,000 to $120,000 · Tier C $68,000 to $100,000. No equity on international offers. No visa sponsorship needed, none implied.

Working from outside the US: core repositories are already owned by engineers outside the US, and you get the same repositories, data, and review authority as anyone else. Reviews are written and asynchronous because the timezone spread forces it, which is the right discipline for regulated code anyway. Structure is contractor or employee via an employer of record, depending on your country. We don't have a two-tier engineering org and we're not building one.

The team: Co-founded by Pranav Rajpurkar (Associate Professor at Harvard Medical School, 150+ publications). MIT and Stanford engineers and fellowship-trained radiologists who read alongside the model.

To apply: Use the application form below. Alongside your CV and GitHub, you'll answer one short question: the last interface you built that someone used for hours a day, and the thing you changed after watching them use it.