Full Stack Developer (Remote)

Who we are: Magellan AI gives brands, agencies and publishers visibility into the podcast advertising market and proof that it works. Our audio analysis and NLP track ad placements across thousands of shows. Our measurement and attribution tools connect those ads to the traffic and conversions they drive, so buyers can optimize campaigns and sellers can prove performance. We are looking to hire an experienced full-stack developer to join our team of seven engineers, to help launch a major new feature and continue to enhance our core product. The job has changed shape over the past year: most of our code is now drafted by AI agents — Claude Code and Codex — working under close supervision. We ship considerably more than we used to, and the scarce skill is no longer typing speed. It is knowing what to build, specifying it precisely, and being able to tell a correct answer from a convincing one. We are seeking someone who can handle many responsibilities, including: Adding new features to our Ruby on Rails application from top-to-bottom, full stack. You will work with operations and sales to help creatively design 'good enough' yet professional solutions.Directing AI coding agents through real work: framing the problem and its constraints, reviewing what comes back line by line, pushing back when a solution has outgrown the problem, and starting over simpler when that is the right call. You own the result regardless of who typed it.Increasing system/code quality by using practices such as TDD and refactoring. Quality around the core of the system has become more important as we scale up — and more important still now that the volume of code we produce has gone up.Writing the design down before the code: specs, architecture decision records, and product documentation for our operations team. Precise writing is a core engineering skill here, because it is both how we align with each other and how we brief our agents.Extending our own tooling — the skills, subagents and CLI helpers that encode how we work. Codifying a workflow so the whole team, and their agents, can repeat it is ordinary, reviewed work here.Building AI-facing product surfaces, like our Ad Intelligence MCP server, where the consumer of the API you design is a language model rather than a browser — with the authentication, policy gates and audit logging that implies.Enhancing usability of existing features via Turbo/Stimulus and other front-end improvements.Troubleshooting and fixing performance issues and complicated logic across Postgres and ClickHouse. Help simplify the system by bringing in smart design decisions.Experimenting and providing stopgap solutions for potential opportunities. Sometimes this involves quick and dirty solutions such as scripting, SQL exports, web scraping and the like.Sharing in providing operational support to make sure the system stays running and address any bugs. (Most usage is during normal business hours...we don't do pagers.)Experimenting via internal hackathons to produce novel solutions and feature enhancements.Software Development at Magellan AI We adhere to a lightweight Agile development process. Every week, we have one formal product meeting with the revenue and operations team to reset priorities. The rest of the time is self-directed, with developers figuring out how to best solve any problems. In general, we are not working against any deadlines and make sure we take the time to do the job right.We practice Collective Code Ownership so that anyone can change any line of code and anyone can introduce new libraries or tools.We deploy code on a regular basis, usually every day. IOW, we ship fast and learn fast. A typical recent week is around 45 pull requests merged.Claude Code and Codex do most of the typing. Engineers spend their time deciding what to build, writing the spec, and judging what comes back.We work one task per git worktree, which lets each of us keep several agent sessions running in parallel on isolated branches. Managing that parallelism is a skill in itself, and we will teach it.Every change runs a review gauntlet before it merges: a local pre-PR review, a cross-agent review where Codex critiques Claude's work, GitHub Copilot's automated review (our branch rules will not let us merge until every thread is resolved), and a human reviewer. Different models catch different things.We invest heavily in the context our agents read: a repository-level conventions file, roughly two dozen project-specific skills, and purpose-built subagents. When someone learns something the hard way, it gets written down where the next agent will find it.Decisions get recorded as ADRs — we are up to 159 — and business rules live in a shared documentation repository linked into every worktree, so both people and agents can find out why the system works the way it does.We usually host a biweekly team retrospective to discuss ways to improve communication and collaboration.We make it easy for the tech-adjacent