Agent Logic Engineer

About the role

ClaimIt runs on a 4-agent system that monitors purchases, drafts claim materials, and prepares them for user approval. The hard work isn't training models — it's prompt design, tool selection, output validation, and recovery from agent mistakes. You'll own that craft. We use Gemini (via Google's Agent Development Kit) with tool routing, output validators, and self-evaluation loops. You'll keep the agents honest, fast, and trustworthy.

What you'll do

• Design and iterate on system prompts, tool descriptions, and few-shot patterns across our 4 sub-agents
• Build validators (structural + semantic) that catch agent mistakes before they reach users
• Implement self-evaluation loops where agents check their own outputs against ground truth
• Design tool interfaces — what tools to expose, when to call them, how to recover from failures
• Run evals end-to-end: build the dataset, score outputs, ship improvements with confidence

What you'll bring

• 2+ years of practical LLM application work (not research; shipped systems serving real users)
• Strong intuition for prompt design and agent debugging — you've spent hours staring at trace logs
• Comfort with eval design: synthetic datasets, regression suites, scoring rubrics
• Strong Python fluency; able to ship infrastructure-adjacent code (FastAPI, async, MongoDB)
• A bias for measurement — 'feels better' doesn't count; show the eval delta

Nice to have

• Experience with Google ADK or similar agent frameworks (LangChain, CrewAI, etc.)
• Background in structured generation (Pydantic models as output schemas, function calling)
• Past work on agents that take real actions in the world (not just chat)