AI Engineer, Research and Development

Every client deployment runs on Prime, our operating system for enterprise AI. Prime captures how a business works, holds shared memory, builds the agents and keeps every action under human control.

This role is about making that stack more capable. You will work at the frontier of agent and model techniques, decide what is worth adopting, and take the ideas that work all the way into the product. The forward deployed engineers are your closest partners: what they meet in the field sets your agenda, and what you build gives them more to deploy.

We are hiring several engineers into this team.

What you will do

• Extend the layers of Prime: capture, memory, agent building, oversight and learning from live operations.
• Design agent architectures for long-horizon work: planning, tool use, sub-agent delegation, state and recovery when a step fails.
• Prototype new model and agent techniques, and measure each one against real client workflows before it goes near production.
• Build the evaluation harnesses, test sets and regression suites that decide whether a change is ready to reach a client.
• Improve how agents remember: retrieval, structured business memory, context management and the limits of long context.
• Make agents faster, cheaper and more reliable: model routing, caching, structured outputs, latency and cost per task.
• Instrument agents end to end, with traces, logs and review queues, so people can see and correct every action.
• Turn needs that keep recurring in client deployments into reusable platform capabilities.
• Follow the research frontier closely: read, reproduce and say plainly what is worth adopting and what is not.

What you bring

• Strong software engineering. Python is your daily language, and you are comfortable in TypeScript.
• Hands-on experience building agents or LLM systems that went beyond a demo and ran for real users.
• Working knowledge of current agent frameworks and SDKs, such as the Claude Agent SDK, the OpenAI Agents SDK, LangGraph or similar, and a clear view of when to skip a framework.
• Experience with tool calling and the Model Context Protocol (MCP), including building servers and clients.
• Experience with retrieval and memory: embeddings, vector stores such as pgvector, hybrid search, reranking and knowledge graphs.
• Experience evaluating LLM systems: LLM-as-judge, human review, offline test sets and online monitoring, with tools such as Inspect, Braintrust, LangSmith or your own.
• Familiarity with how current frontier models behave, including reasoning models, structured outputs and computer use, and where each one breaks.
• A scientific habit: you state a hypothesis, design the test and change your mind when the result says so.
• Experience taking research ideas into production code, with solid fundamentals in Postgres, queues, containers and cloud infrastructure.
• Clear writing and the confidence to explain technical trade-offs to non-specialists.

Strong advantages

• Experience with post-training: supervised fine-tuning, preference optimization or reinforcement learning with verifiable rewards, using tools such as TRL, Unsloth or similar.
• Experience serving open-weight models with vLLM, SGLang or similar, and judgment about when they beat a hosted model.
• Experience with durable workflow engines such as Temporal for long-running agents.
• Experience with observability for AI systems: OpenTelemetry, tracing and cost tracking.
• Experience with voice and real-time agents, using tools such as LiveKit or the OpenAI Realtime API.
• Published research, open-source work or benchmark results.

How we hire

We keep the process short: an introductory conversation, a practical exercise based on real work, and time with the people you would work alongside.

About Meridian

Meridian builds and operates AI agents for businesses that run on people and process. We map how the work gets done, deploy agents that complete it end to end under human supervision, and stay to run them in production.