Software Engineer, AI

Bilt is the membership for where you live. We started by rewarding people on their biggest expense, rent, and have grown into a network connecting members, property managers, and local merchants across dining, travel, fitness, and everyday spending.

Concierge runs in the Bilt app, on the web, and over SMS, iMessage, and email. Members use it for dining, rent, payments, travel, and support. Restaurants and property managers use it to talk with their customers, and their staff use it as an assistant. It handles tens of thousands of conversations a day, with 2–3x spikes at the start of every month, and you'll build and run the backend behind all of it.

• Agent capabilities end to end: tools, workflows, prompts, and the integrations with Bilt and partner systems behind them
• The agent runtime: multi-agent orchestration, context and memory, prompt caching, and routing across frontier and open-weight models
• Performance: p50 and p99 across the full request path, bursty traffic, and cost per conversation
• Eval harnesses, LLM judges, and conversation replay, so every model and prompt change is measured before it ships
• Assistants for merchants' staff that work from the business's own data, and browser agents that can act where no API exists
• Proactive features that drive engagement with members and merchants' customers: briefings, follow-ups, and outreach that land at the right moment
• The conversation platform: one inbox where the agent, merchant staff, and Bilt support share threads across SMS, iMessage, email, web, and in-app chat
• Messaging: realtime streaming replies, attachments, message ordering and idempotent delivery, carrier and iMessage sending limits, and handing off to a person at the right moment
• Authentication and fine-grained authorization, so the agent only sees and does what it should for members, merchants, and staff

We are an AI-native team. Agents are part of how we write, review, debug, and operate everything.

• Writing. Coding agents like Claude Code and Devin write much of our code. You design it, review it, and own what ships
• Reviewing. Every PR gets AI review against rules we write down (prompt-cache pitfalls, security, performance) before a human codeowner signs off
• Debugging. Agents have access to our production tooling through an internal MCP gateway: logs, traces, analytics, feature flags, and the service catalog. Investigations usually start with an agent pulling the evidence
• Sharing context. Repos carry agent instructions, and we keep a shared library of skills and playbooks, so recurring work becomes a single command and every agent starts with the team's context
• Owning. Each engineer owns a large surface from design through rollout and on-call. We ship daily behind flags, treat performance and availability as acceptance criteria, and read real conversations, not only dashboards

• 4+ years building and operating backend services in production
• Strength in at least one backend language and experience running services in the cloud. We mostly use Python (FastAPI), Java (Quarkus), and Go on Google Cloud
• Solid distributed systems fundamentals (async processing, pub/sub, idempotency, retries, caching), and you debug across services with logs, traces, and metrics
• Hands‑on experience with LLMs, at work or on your own, and curiosity about making them reliable
• Product judgment: you can read a bad conversation and tell whether the cause is the prompt, a tool, the data, or a service
• Daily use of AI coding agents, and good judgment about when to trust their output
• Clear writing. Design docs, PR descriptions, and the prompts you give agents all matter here

• Shipped an LLM-powered feature to real users, and can explain what broke and how you…