Who we are:

Position overview:

What you'll do:

• Design and build agentic workflows on top of MIA's MCP server - tool definitions, complex reasoning and multi-step planning, context/session management, and error recovery for a system real customers query daily.
• Own the harness quality bar: work with frontier models (Gemini, Claude and others) to close the gap between "technically works" and "production-grade agent behavior". Prompt/Skills/system design, reflection steps, structured tool responses, and graceful degradation when tools fail or return partial data.
• Build and maintain evaluation pipelines for agent quality. Accuracy, tool-call correctness, hallucination/fabrication detection, and regression testing across our causal models and workspaces.
• Extend the MCP surface: design new tools, metadata extensions, and (where relevant) interactive MCP App / UI resources; so agent output isn't just text but something a user can act on.
• Work with structured marketing data (BigQuery/Spanner-backed outputs and translate it into agent-consumable context and user facing narratives based on the user's persona.
• Debug production agent issues systematically: stuck sessions, workspace binding failures, inconsistent tool outputs, model-vs-model disagreement and turn recurring failure patterns into fixes, not one-off patches.
• Stay current on the agentic AI and MCP ecosystem and bring back what's actually usable,  new protocol capabilities, orchestration patterns, eval techniques; rather than chasing every new framework.

You're a strong fit if you:

• Have 3–6 years of experience in ML/AI engineering, with recent hands-on work building and shipping agentic systems (not just RAG chatbots) - ideally something a real user depended on.
• Have built or deeply worked with MCP (Model Context Protocol), or a comparable tool-calling / function-calling framework, and understand the practical failure modes of agent-tool interaction.
• Are strong in Python, comfortable with LLM orchestration (whether via raw API tool-use, LangGraph, Claude Agent SDK, or your own harness), and know when NOT to reach for a framework.
• Have production experience with evals - you can articulate how you'd measure whether an agent got better or worse after a prompt or tool change, and you've actually built that measurement, not just read about it.
• Are comfortable owning ambiguity: given a business use case (e.g. "help a marketer trust a budget reallocation"), you can scope it, prototype it, and ship a v1 without needing a fully-specified spec.
• Have cloud deployment experience.

Nice to have:

• Familiarity with MCP Apps / interactive UI resources, or building agent-facing frontends (we use TanStack Query/Router, Zustand).
• Experience with vector databases and RAG, even if this role is more agentic-orchestration than retrieval-heavy.
• A track record of 0 to 1 builds; you've taken something from "nobody's built this yet" to "customers use this every day."

Our stack:

What's in it for you:

• Ground-floor ownership of the agentic layer at one of the fastest-growing MarTech companies right now, your decisions on harness design and tool architecture ship to production customers within weeks, not quarters.
• Small, non-bureaucratic team with real empowerment; you'll work directly with the CTO and founding engineers, not through three layers of process.
• A genuinely hard, well-scoped problem: agentic reasoning over causal marketing models, where correctness actually matters to the customer's budget decisions.
• Competitive compensation and benefits, and a highly profitable, growing organization with real room to accelerate your career.
• A team that bonds over tea, movies, and Friday hangouts and takes work-life balance seriously.