Staff Machine Learning Operations Engineer - Computer Vision

ATI•Woburn, MA

11h

About The Position

BrakeWise is our production brake inspection product: a mobile application paired with a camera probe that technicians use to assess pad and rotor condition during live service work. The machine learning behind it is a multi-stage pipeline of segmentation and classification models that turn raw imagery into a wear assessment a shop can act on and charge for. That pipeline works, and it is an MVP. It runs on Cloud Functions, and it will not carry us to the customer volume we're signing. We're looking for a Staff MLOps Engineer to own it — to take it from a working prototype to a serving architecture that holds up under real throughput, with the latency, cost, and reliability characteristics a paying customer expects. You'll own every aspect of how our models reach production and how they get better: serving infrastructure, deployment and rollback, monitoring and drift detection, the retraining loop, and the evaluation discipline that tells us whether a new model is actually an improvement. Model accuracy here has commercial consequences — a bad wear call is either a missed repair or an unnecessary one, in front of a customer. This is also the senior cloud architecture voice on the team. You'll partner closely with our Staff Full Stack Engineer, who owns the mobile app and customer dashboard, reviewing designs and setting GCP practices across the platform rather than only within the ML stack.

Requirements

8+ years of professional engineering experience, including several years owning machine learning systems in production — not solely model development
Demonstrated experience taking a computer vision pipeline from prototype to production scale, serving real users at meaningful volume
Deep experience deploying and operating segmentation and classification models, including multi-stage pipelines where one model’s output feeds the next
Strong cloud infrastructure background, preferably GCP — Vertex AI, Cloud Run, GKE, Cloud Functions, Cloud SQL, Pub/Sub, and Docker
Production-grade Python, and fluency with PyTorch or TensorFlow
Hands-on experience with model serving and optimization — Triton, TorchServe, ONNX, TensorRT, quantization, or equivalent
Experience owning deployment and support for a live system, including incident response, rollback, and on-call
Experience building data and labeling pipelines with dataset versioning and reproducible evaluation
Comfort with infrastructure as code (e.g., Terraform) and CI/CD automation (e.g., GitHub Actions)
Sound judgment on the accuracy, latency, and cost trade-offs that determine whether an ML product is viable
Excellent problem-solving, debugging, and communication skills, including with non-technical stakeholders

Nice To Haves

On-device or edge inference experience (Core ML, TensorFlow Lite, ExecuTorch) and integration into mobile applications
Active learning or human-in-the-loop labeling systems
Computer vision on small, long-tail, or industrial inspection datasets rather than large public benchmarks
Experience with camera and sensor integration, or working alongside hardware teams on capture quality
Experience with robotics, IoT, or edge computing (ROS or similar platforms)
Familiarity with automotive service, dealership operations, or DMS ecosystems
Contributions to open-source projects

Responsibilities

Own the multi-stage inference pipeline (segmentors and classifiers) end to end — serving architecture, latency, throughput, reliability, and cost per inspection
Re-architect the pipeline off its current Cloud Functions MVP onto infrastructure that scales: containerized inference, GPU-backed or accelerated serving where it pays for itself, queueing, batching, and autoscaling
Own model deployment: versioning, staged rollout, canary and shadow evaluation, and fast rollback when a model regresses
Build and own the improvement loop — field data collection, labeling workflows, dataset versioning, evaluation harnesses, and regression suites that catch quality loss before customers do
Monitor model quality in production: drift detection, segmented performance analysis, and triage of real-world failures against real inspection imagery
Define the metrics that matter commercially — false-positive and false-negative rates on a wear call, technician override rate, unit inference cost — and report against them
Improve model performance directly: architecture selection, augmentation, hard-example mining, and quantization or distillation where latency and cost demand it
Evaluate on-device versus cloud inference trade-offs for the mobile app, and own whichever path we choose
Establish MLOps foundations: reproducible training, experiment tracking, CI/CD for models, and infrastructure as code
Serve as the cloud architecture counterpart to the Staff Full Stack Engineer — reviewing designs, setting GCP best practices, and raising the platform’s infrastructure bar
Work with hardware and field operations on capture quality — lighting, focus, and probe positioning — since upstream image quality sets the ceiling on model performance
Own production support for the ML stack, including incident response and on-call participation for inference availability
Proactively identify technical risks and architectural trade-offs, and communicate them clearly to leadership