Technical Case: Clinical Operations Assistant

Overview

Build a small agentic "clinical operations assistant" that:

• interprets user intent,
• plans and executes multi-step tool calls,
• answers coordinator questions from database-backed tool results,
• and proposes actions that require human approval before execution.

The provided database.db (SQLite) is the only data source you should use, containing a couple of tables with synthetic clinical data.

• Do not change the schema of the database, use as is.
• You are expected to both read and write rows in the database.
• During review, we may swap in a different .db file with the same schema to verify robustness.

In order for this case to be deterministic, assume current date/time is always 2025-03-16T09:00:00Z.

You may use any language/framework and any helper tools, including AI coding assistants and fully LLM-generated code.

An API key for the LLM used by your solution will be provided. Details on how it is shared will be provided to shortlisted candidates.

Functional Requirements

Authentication and authorization are explicitly out of scope for this case. You are not expected to implement login, user/session management, RBAC/permissions, OAuth, end-user API key management, or any other auth mechanism. Integrating a provided LLM API key into the runtime configuration of your solution is sufficient.

1. Implement an agentic interactive chat interface that can be demonstrated live during review (CLI, web interface, or equivalent).

• The interface must support multi-turn follow-up questions in the same session.
• The interface must support interactive approval decisions (approve / deny) for pending mutating actions.

2. Implement tool-based retrieval:

• All data must be fetched through tool abstractions.
• Any mutating action (write/update/delete) must require HITL (Human-in-the-Loop) approval before execution.
• Read-only operations can execute directly (no HITL required).
• Provide transparency over which tools were called, with which parameters, and what results were returned.
• Pending actions should remain addressable in the live interaction so a reviewer can approve/deny them in subsequent turns.

3. Add focused tests where appropriate (unit and/or integration)

Example queries which should be supported:

• "Which patients are scheduled for appointments in the next 7 days?"
• "Who has abnormal (HIGH) lab results in the last 14 days?"
• "Create an appointment for Patricia Adams with Dr. Alice Nguyen next week as a follow-up."

Expected Deliverables

Source code, documentation and example inputs/outputs demonstrating the system's capabilities.