Agent Requirements

Agent Objective:

To serve as a "Digital Residency Coordinator" that helps users navigate the Medtrics platform, troubleshoot login issues, and provide "how-to" guidance for modules like Curriculum Mapping, Evaluations.

Autonomy & Constraints:

Autonomy: The agent can search the Medtrics Knowledge Base and provide direct links to tutorial videos. It can check if a user's account is active via API.

Constraints: It cannot change grades, modify clinical schedules, or access sensitive IRIS cost report data without human administrator approval. It must adhere to FERPA and HIPAA data privacy standards.

Tool & Action Design

Knowledge_Base_Search: A RAG-based/Vectoreless tool to search your documentation, tutorial transcripts.

User_Status_Check: Verify the User

Ticket_Creator: If the agent can't solve the issue, it uses this tool to create a Optimus and notify on slack

Auto_Step_Assist: To assist user to how to perform certain task(future scope)

Agent Architecture

Orchestrator : Acts as the brain. It receives the user query and classifies the intent

Specialized Workers:

Support Worker: Expert in SSO, browser issues, and mobile app syncing.

flowchart TD

Client["Client User - Persona and SSO Session"]

Orchestrator["Orchestrator Agent - Multi-Agent Brain - ReAct and CoT Planning - Intent Decision - Escalation Handoff"]

Memory["Conversation State and Memory Buffer - Short-term sliding window - Contextual Metadata"]

TechAgent["Technical Support Worker Agent - Login SSO App Sync - SSO Status API - Zendesk Jira Handoff"]

WorkflowAgent["Workflow Guidance Worker Agent - MSPE Evaluations Duty Hours - RAG Search - Module Deep-Link"]

VectorDB[("Vector Database - RAG documentation")]

ExtAPIs["External Services and APIs - MedTrics Platform"]

CoreDB[("Core Platform Database - User and System Data")]

Output["Output Generation and Safety Guardrails - LLM Polish - FERPA HIPAA - RBAC - PII Masking"]

Client --> Orchestrator

Orchestrator <--> Memory

Orchestrator --> TechAgent

Orchestrator --> WorkflowAgent

WorkflowAgent --> VectorDB

TechAgent --> ExtAPIs

WorkflowAgent --> ExtAPIs

VectorDB --> ExtAPIs

ExtAPIs --> CoreDB

TechAgent --> Output

ExtAPIs --> Output

CoreDB --> Output

Output -->|Final Filtered Response| Client

Memory & Context Design

Long-term : Uses RAG .

All MedTrics documentation, FAQ pages, and transcripts are indexed in a Vector Store (like Pinecone).

Context Management: Upon login, we inject Metadata into the system prompt like user details and institute details.

Core Reasoning

Planning: here is the flow of agent process.

User: "My student's evaluation is missing."

Agent Plan: 1. Identify student. 2. Check if evaluation was sent. 3. Verify if it's in "Draft" or "Pending" state.

Self-Correction: Before answering, the agent performs a Verification Loop: "Does this answer contradict the permission issue?"

Human In Loop:

If the confidence score for a response is below 85%, or if an API call fails, the agent triggers a "Human-in-the-loop" handoff to the actual Medtrics support desk.

If user not satisfied with answer then pass on to human.

Any critical issue reported regarding USER , Scheduling,IRIS directly transfer to Human agent.

Safety & Permissions

Permission Boundaries: The agent inherits the user’s platform permissions.

Content Filtering: We use Guardrails to ensure the agent never provides medical advice, as it is a management tool, not a clinical tool.

PII Protection: Any output containing Social Security Numbers or sensitive health data (if applicable to IRIS reporting) is automatically redacted before being shown to the user.