Full Stack AI Engineer

This role is suitable for a Senior Full Stack AI Engineer with 5+ years of experience owning and shipping production software end to end.

The ideal candidate has strong Python and TypeScript skills, experience with React, backend services, databases, AWS infrastructure, LLM agents, tool use, retrieval, structured outputs, evaluations, and AI coding-agent harnesses such as MCP servers, skills, hooks, and subagents.

They should be comfortable turning ambiguous healthcare-operations problems into products, designing APIs and data models, building web and mobile features, deploying and monitoring services, handling PHI securely, and managing production incidents independently.

How would you design an AI agent for insurance benefits verification, including tool use and human escalation?

I would define structured inputs and outputs, expose controlled tools for payer lookups and document retrieval, and require the agent to cite evidence for each result. Low-confidence or conflicting cases would be routed to an operator through a durable workflow, with all decisions and overrides auditable.

How would you evaluate an LLM-powered workflow used for prior authorization or clinical documentation?

I would create a representative, de-identified evaluation set with expected structured outputs, tool-call behavior, safety constraints, and escalation rules. I would track accuracy, unsupported-claim rate, latency, cost, completion rate, and human override rate both offline and in production.

How would you build an agent harness using Claude or OpenAI models, MCP servers, skills, and hooks?

I would define project instructions and reusable skills, expose narrowly scoped MCP tools with explicit permissions, and use hooks for validation, formatting, and security checks. Subagents could handle independent tasks, while a review step would verify changes, tests, tool usage, and sensitive-data handling before integration.

How would you protect PHI in a Python or TypeScript healthcare service?

I would enforce least-privilege access, encrypt data in transit and at rest, isolate environments, and avoid placing PHI in application logs, prompts, traces, or error messages unless explicitly required and controlled. Access, mutations, exports, and administrative actions should be auditable and monitored.

How would you model and expose a new clinic-operations feature across PostgreSQL, APIs, and the frontend?

I would start with the workflow and invariants, then design normalized PostgreSQL tables, migrations, indexes, and explicit state transitions. I would expose a versioned API or GraphQL contract with authorization at the resolver or service layer, then build the React or React Native UI around loading, validation, failure, and optimistic-update states.

When would you use Temporal for a healthcare workflow?

Temporal is appropriate for long-running, retryable, multi-step processes such as benefits verification or prior authorization, especially when they include timers, external callbacks, and human review. Activities should be idempotent, workflow state should remain deterministic, and retries should distinguish transient failures from business failures.

How would you design reliable structured outputs from an LLM agent?

I would use a strict schema with enums, required fields, and validation at the application boundary, preferably through provider-supported structured generation. Invalid or incomplete responses would be repaired or retried with bounded attempts, while high-risk fields would require evidence, deterministic checks, or human approval.

How would you deploy and operate a new service on AWS ECS Fargate?

I would package the service in a reproducible container, provision infrastructure with Terraform, and deploy through GitHub Actions using separate environments and controlled secrets. Datadog dashboards and alerts should cover latency, errors, saturation, workflow outcomes, dependency failures, and business metrics, with a documented rollback and incident process.

How would you use pgvector and retrieval in an AI feature?

I would chunk and embed authorized documents with metadata such as tenant, patient, document type, and effective date, then apply strict metadata filtering before similarity search. Retrieval quality should be evaluated with recall and precision metrics, and the agent should cite retrieved sources rather than treating vector results as unquestionable truth.

How would you safely use Claude Code or Codex across a large existing codebase?

I would provide repository instructions, architecture guidance, test commands, coding conventions, and security rules, then connect only approved tools through MCP. I would split work into small tasks, run agents in parallel where dependencies allow, and personally review diffs, tests, migrations, permissions, and observability before merging.

3Y Health is building an AI-native operating system for healthcare practices, combining software with managed back-office operations for clinics. As a remote contractor based in Argentina, you’ll own end-to-end features—from AI agents for insurance, intake, claims, and clinical documentation to clinician-facing web and mobile products—while designing services, infrastructure, evaluations, and production monitoring. The stack includes Python, TypeScript, Kotlin, React, Next.js, React Native, PostgreSQL, Temporal, AWS, Terraform, and Claude/OpenAI models, with daily use of Claude Code and Codex. Salary is not specified, but this role is worth applying for if you want high ownership, meaningful healthcare impact, and the opportunity to shape advanced AI engineering practices from the ground up.