AI Engineer; Agent Platform

Location:

Hybrid (Amsterdam Office) | Salary: €70K – €110K

Bonnie is the AI assistant that answers the phone for local businesses — restaurants, clinics, salons — across phone, Whats App and chat. 500+ businesses in 14 countries use us today. Millions of conversations, each one a real customer who expects an answer now.

Engineering is small and broad: a handful of engineers across data, APIs, Dev Ops, AI agents and cloud, plus a designer. Everyone ships to production, and everyone talks to the businesses using Bonnie — the loop from a change to a restaurant owner telling you what they make of it runs in days, not quarters.

You own the decision layer: the tool-sets the assistant calls mid-conversation, the orchestration and routing across models, the configuration platform that lets every business behave differently, and the evaluation system that keeps all of it reliable at millions of conversations.

We are hiring multiple people into this scope — one senior, one medior — so the depth we expect scales with the level you come in at.

The stack: Python and FastAPI, LiteLLM for model routing, language tools (Lang Fuse, Langsmith) for tracing and evaluation, MySQL and Postgres, queues and workers, Docker. Claude Code, Codex and Cursor in the daily loop.

Your first project is the evaluation harness: replaying real conversations against a change and scoring what comes out. It is the fastest way to learn the product — within a month you will have touched the call path, the tool-sets and the data behind them.

• Design and own Bonnie's tool-sets — availability, reservations, modifications, escalation — including their contracts, failure behaviour and latency budgets.
• Build the evaluation system. Replay real conversations, score outcomes, catch regressions before they ship. This is how Bonnie stays reliable as volume grows, and it is the highest-leverage work on the team.
• Build the configuration platform: a schema-driven model built from blocks for per-business behaviour — versioned, validated, and safe to change without a deploy, so a change of tone or opening hours never needs an engineer.
• Orchestrate the models. Routing, fallbacks, guardrails, and the cost and latency trade-offs between them.
• Work on the call path. Latency budgets, turn-taking and interruptions, transfers, SIP, load balancing across STT and TTS providers. A hundred milliseconds is audible, and this is where the agent meets the phone line.
• Run what you build. Deploys, observability, and a two-week support shift that comes round about once every two months. Outside office hours is light — historically a few hours on an incidental Saturday, roughly once a quarter. Mostly we keep an eye because we built Bonnie and we care how it behaves, not because a rota says so.

• 3+ years building and operating production backends
  , at least one of them on LLM-powered systems real users depended on. For the senior role: 5+ years, and a system you designed rather than contributed to.
• Agentic systems, and where they break. Tool calling, multi-step orchestration, structured outputs, retries, partial failure, context management. These are distributed systems problems with a new failure mode.
• A point of view on evaluation. You have had to prove an AI system got better, and you know why a good demo is not proof.
• Async Python and API design. Contracts, idempotency, timeouts and failure modes before the handler. Comfortable in a relational database and a queue.
• AI-native, with accountability. Coding agents are in our daily loop. Use them, and explain, test and own what they produce.
• Comfort in a codebase you did not write. Ours got us to 500+ customers at speed, and it shows. We are carving it into clear domains while…