Computer Vision SDE II, Scanless Technologies

Description

Our team builds the technology that allows Amazon's fulfillment systems to see and understand their environments without requiring employees to manually scan every item. We develop and deploy camera-based solutions inside robotic and human-operated workstations that can automatically identify, count, and track items and packages at high speed and high accuracy. When a new type of workstation needs this capability, our team builds it end to end: selecting and qualifying the right cameras, writing the software that processes what those cameras see, and supporting the system once it is running across many sites.

You will design and build the software that runs on edge computing devices installed inside fulfillment workstations. This software connects to various camera systems, processes images in real time, and makes decisions that previously required manual effort. You will ensure these systems work reliably across hundreds of locations where lighting, hardware configurations, and operational demands vary. Your solutions will directly reduce package handling errors and lower equipment costs by enabling simpler, more capable hardware setups to replace complex multi-camera arrangements.

This role requires both strong engineering skills and a willingness to invent. You will find opportunities to bring our vision technology into new types of workstations and fulfillment scenarios, taking ideas from early testing through to full-scale production deployment. You will build the automated testing and release systems that let the team deliver updates confidently, and you will create tools that allow partner teams to set up new deployments independently.

Your work will directly shape the cost, reliability, and reach of autonomous package recognition across Amazon's fulfillment network.

Want to see our technology in action? Our team built the vision system now running inside thousands of Amazon delivery vans that instantly identifies the right package for each stop using projected light, no scanning required. See how it works:

Key job responsibilities

- Evaluate, prototype, and deploy novel computer vision capabilities that reduce hardware costs and improve item identification accuracy

- Design and build edge inference and deployment infrastructure that moves computer vision models from experimentation to fleet-wide production across fulfillment workcells

- Develop and extend the software stack running on edge computing devices, integrating first-party and third-party sensors into production systems that meet strict reliability and accuracy targets

- Build sensor qualification, calibration, and validation tooling that enables new camera hardware to be onboarded and deployed with minimal manual effort

- Own data collection and monitoring pipelines that capture production telemetry, support model retraining, and drive measurable defect reduction across deployed stations

- Create self-service tooling and workflows that enable partner teams to onboard new workcell deployments without dedicated engineering support

- Design fleet-wide observability systems: health monitoring, sensor anomaly detection, performance alerting, and proactive diagnostics that sustain high availability targets across hundreds of stations

- Define and implement hardware abstraction interfaces for new sensor systems, ensuring every product team benefits from newly qualified hardware without code changes

- Drive operational excellence: automated test and release pipelines, incident response, runbook development, and reliability improvements that sustain uptime targets at scale

A day in the life

We build the systems that give Amazon's fulfillment network the power of sight. Our team turns computer vision research into production reality, deploying autonomous perception capabilities to robotic and human-operated workcells that process millions of packages daily. As an SDE on this team, you own the full path from sensor to decision: the software running on edge devices, the inference pipelines serving models in real time, and the tooling that lets new workcell types go from concept to deployment at speed.

On any given day, you might be writing the abstraction layer for a newly qualified camera system so every downstream team gains its capabilities instantly. You could be shipping an automated calibration workflow to dozens of fulfillment sites, diagnosing why a model's accuracy drifted at a specific station, or building a fleet-wide health monitoring system that catches sensor degradation before operators notice it.

You will prototype novel sensing approaches that replace expensive multi-camera setups with smarter, cheaper alternatives. You will design self-service tooling that lets partner teams deploy perception capabilities without waiting on your team. You will push code that runs at the physical edge of Amazon's network, inside the workcells where robots and humans move real inventory at real speed. The problems are tangible, the feedback loop is…