Senior Software Engineer - Applied AI

Senior Software Engineer - Applied AI

SuperClaims · Health insurance claims

About the role

We're looking for a Senior Applied AI Engineer with startup experience who can own AI products end to end — from designing agentic workflows to shipping production systems.

You'll build AI-powered health insurance claim adjudication workflows that turn complex documents into explainable, reviewable decisions. This includes document intelligence, tariff analysis, deductions, multi-agent systems, and human-in-the-loop validation.

What you own

• The claim pipeline from document intake through adjudication: extraction, tariff matching, deduction logic, and the human-in-the-loop step.
• Agent workflows in code (LangGraph or equivalent): graph state, tool calls, structured outputs, retries, and a stop condition when the model should hand off.
• The APIs and background workers that run those workflows in production (long jobs, heartbeats, idempotency).
• How a model is allowed to act: which principal can call which model, what the output schema is, and how a reviewer can reverse a single AI decision without re-running the world.
• Evaluation of the workflow on real claim shapes — not a leaderboard score. You can show where it fails and what you changed.

Our stack

Product: Python, FastAPI, LangGraph, Pydantic.

Workers: Temporal, Celery, or RabbitMQ — long-running claim jobs, not request/response only.

Data: PostgreSQL via PgBouncer. Object storage on S3 (claim files and reference documents stay private).

Models: Vertex AI (Gemini, including models that expose a thinking-level control) and Amazon Bedrock. Called from our code, least privilege, no shared keys.

Runtime: Docker locally. ECS, ALB, IAM, VPC, and a secrets store on AWS. Kubernetes is not required.

Requirements

• 3–5 years building backend product code in Python, including APIs other teams or reviewers depend on.
• Shipped an LLM workflow in production code, not a chat UI: structured output, tools, and an explicit path for "the model should not decide this." LangGraph strongly preferred; a comparable graph or agent framework is fine if you can explain the state and the handoff.
• A workflow you personally built that called a hosted model from your code (Vertex, Bedrock, or similar). You can name the project, the model, and how credentials were scoped.
• Multi-step background work in production: retries, idempotency, and a human step. Temporal, Celery, or RabbitMQ all count. We will not drop you for not having used one of the three.
• PostgreSQL used with transactions, migrations, and a connection pooler. You do not "fix" pooling by bypassing it.
• File and object-storage flows (S3 or equivalent) where private documents stay private.
• Enough AWS to deploy and debug your own service: containers (ECS or similar), IAM roles, security groups, secrets. You can read a task role. You do not need to own the account.
• You treat model output as untrusted input. Schema checks, allow-lists, and a reversible human override are part of the design, not a follow-up ticket.
• AI coding tools are welcome. You review the diff and can defend it in the PR.
• Written English clear enough for a reviewer note, a short design doc, and a PR description. Claims-domain language can be taught. Careful handling of claim documents is not optional.