Senior Software Engineer - Mercury Command

What you'll do

Ship new capabilities users love:

• Design and ship new Command skills, the domain-specific instruction sets that teach the model how to handle workflows like sending money, managing invoices, and understanding cash flow
• Design and build agentic workflows in Command, defining the architecture for how multi-step agent interactions should work as we extend what the product can do on a customer's behalf
• Work with backend teams to define tool schemas for new capabilities, shaping the data contracts between Mercury's business logic and the model
• Own new capabilities end to end, from the system prompt to the frontend component that renders the response

Own the LLM layer:

• Maintain and evolve Command's prompt architecture: the system prompt, skill loading system, session context, and the policy and compliance layers underneath
• Tune model behavior: reasoning effort, prompt caching strategy, fallback chains, and the streaming patterns that make the product feel fast
• Stay current with how models are evolving and bring that knowledge back to how Command is built

Build quality in:

• Write and expand Command's eval harness, adding cases that cover new capabilities and scoring rubrics that detect regressions before users do
• Partner with product and compliance teams to define what "working correctly" means for each new capability, then build the tests that prove it
• Own the reliability and quality of what you ship, from initial design through post-launch monitoring

The ideal candidate

• Has 7 or more years of software engineering experience, with deep technical expertise building and scaling LLM-powered applications in production
• Has gone beyond shipping a first version: you have scaled an LLM-powered product, dealt with the reliability and performance problems that come with real usage, and made it better over time
• Has experience designing agentic systems and has opinions about how to architect multi-step workflows that are reliable, explainable, and safe to run on behalf of real users
• Has built eval infrastructure and can write cases that actually measure whether the product works, not just whether the model outputs something plausible
• Understands the real tradeoffs in LLM deployments: latency, cost, compliance, and what breaks in production that doesn't show up in demos
• Has opinions about what makes an AI product trustworthy, not just impressive, and can build toward that bar
• Is comfortable with TypeScript and willing to learn Haskell for