Software Engineer, AI for the Planet

Ai2•Seattle, WA

1d•Onsite

About The Position

We are a small engineering team at the Allen Institute for AI working on AI for the Planet. We're working on maritime conservation, food security, disaster resilience, and climate solutions with some of the most impactful organizations on the planet. We work very closely alongside a ML research team and our Product & Partnerships teams, focused on building products that support our environmental and high-impact users. Today, our team works on two products: Skylight uses AI to detect illegal, unreported, and unregulated fishing in real time. Governments, enforcement agencies, and conservation organizations in 95+ countries use it to protect their waters. Our advanced AI-powered platform delivers real-time vessel detections and actionable insights that empower enforcement agencies globally to protect marine ecosystems. Read more at https://allenai.org/skylight . OlmoEarth is an open, end-to-end platform built around our family of foundation models for Earth observation. The OlmoEarth Platform enables our users to create custom fine-tuned models to detect and classify novel geospatial features. The platform handles the full loop: imagery acquisition from Sentinel-1, Sentinel-2, and Landsat; annotation; distributed training and inference; and a viewer so the outputs are usable by people who aren't ML experts. Partners today include NASA JPL (wildfire risk), IFPRI (crop mapping in Kenya), Global Mangrove Watch, and the Amazon Conservation Alliance. Read more at https://allenai.org/olmoearth .

Requirements

5+ years of professional software engineering experience in industry. Internships, graduate school, and research positions are valuable but do not count toward this.
Experience working as a generalist software engineer across multiple parts of the stack and product lifecycle.
Strong foundations in web applications, data pipelines, distributed systems, and modern cloud tooling.
The specific frameworks matter less than demonstrated ability to pick up new technologies as the field evolves.
You're using modern AI tooling (e.g. Claude Code, agentic workflows) to move faster and rethink how engineering gets done.
A track record of taking software products end to end. Shaping requirements, designing architecture, shipping to external users, and continuing to develop them over time based on user feedback.
Ownership over production systems. Including taking on-call rotations, troubleshooting production issues, and digging into logs, metrics, and code to develop real, actionable insight when something needs attention.
You write clear technical plans, give and receive feedback well, constantly prioritize, and can guide stakeholders through the details.

Nice To Haves

Experience at small or growth-stage companies, where you own outcomes end to end without heavy process scaffolding.
Hands-on experience integrating machine learning models into production systems: deployment, monitoring, scaling real-time inference, and iteration.
You collaborated directly with researchers to bring models out of a research context, and you built user-facing applications where AI outputs need to be communicated clearly to non-technical users.
You're opinionated about software engineering practice: coding patterns, breaking down work, code review, testing, build systems. You bring that judgment to the team while staying open to other perspectives.
A demonstrated track record of technical depth and self-directed learning - for example: open-source contributions, technical writing, conference talks, sustained side projects.
Open to occasional international travel to meet directly with the people using what we build.

Responsibilities

Automated model development: We're building the OlmoEarth Platform to enable users to go from raw tabular data to a fine-tuned, evaluated, production-ready computer vision model, without needing an ML engineer. The underlying infrastructure allows us to run jobs across thousands of parallel GPUs and terabytes of satellite imagery - covering continent-sized areas for fractions of a penny per square kilometer. We're also pushing into agentic approaches: agents that help with dataset discovery, preparation, and augmentation, and agents that explore model configurations and architectures to find the right setup for a given use case.
Deploy multi-tenant agents: We are building a multi-tenant agent-orchestration platform to power Skylight's next generation of AI products - starting with Shippy, our maritime-domain-awareness agent. Every end user gets their own isolated sandbox: a per-user container stack with persistent GCS-backed state, a conversation history, and a hardened network boundary where the agent runtime can run free, in a secure environment. This platform will allow us to launch agentic-powered products without re-inventing the wheel every time.
Sentinel-2 vessel detections: We use the Sentinel-2 Satellites from the European Space Agency to detect locations of vessels globally in near-real-time. Our data-pipelines download imagery as soon as it’s available and run our state-of-the-art computer vision models to detect vessels and make these observations available to our users, typically in under 4 hours from an image being taken. You can read more about this project in our blog post here .