Deploy & Connect

From Localhost to Production

A working HR assistant on localhost is a proof of concept. Deploying it means making it accessible, reliable, and auditable. For HR systems, this last requirement — auditability — is non-negotiable. Every query and response must be logged for compliance purposes, guardrail triggers must be monitored, and data freshness must be tracked.

Production Deployment

Environment Management

An HR assistant has more environment variables than a typical web app because of the sensitivity of the data:

Never store these in code. Use Vercel environment variables (encrypted at rest), with separate values for development, preview, and production.

Health Checks

The /api/health endpoint verifies:

If data freshness fails, the system should still serve cached responses but display a warning: "Policy data may not reflect the latest updates."

Slack Integration

Most employees don't want to open a separate app for HR questions. They want to ask from where they already work — Slack.

Slash Command

/hr What's our parental leave policy in California?

The response appears as a rich Slack message with:

@Mention

@HRAssistant How do I file an expense report for international travel?

Responds in the thread, maintaining conversational context for follow-up questions.

Ephemeral Responses

For sensitive queries, the response is ephemeral — only visible to the person who asked. This is critical for:

The query classification from Module 4 determines whether a response should be ephemeral. Queries tagged as sensitivity: "sensitive" or sensitivity: "restricted" always get ephemeral responses.

Channel Restrictions

The HR bot only responds in approved channels. This prevents scenarios where an employee asks a sensitive question in #general and gets a public response. Approved channels are configured in the bot's settings:

MCP Server

The Model Context Protocol (MCP) server exposes your HR knowledge base as a data source for other AI tools. When an employee uses Claude Desktop or Cursor, they can query HR policies directly through the MCP connection.

Tools

Resources

Why MCP Matters

MCP turns your HR knowledge base into a universal data source. An engineer using Claude Desktop can ask "Does our data security policy allow using external AI tools?" and get an answer sourced from your actual policy documents — without opening the HR app. The same guardrails apply: no salary disclosure, no legal advice, confidentiality enforcement.

Monitoring & Compliance

Real-Time Dashboard

The monitoring dashboard tracks:

┌─────────────────────────────────────────────────┐
│  HR Assistant Dashboard                          │
├──────────────┬──────────────┬───────────────────┤
│  Queries/day │  Avg latency │  Cache hit rate    │
│     247      │    1.2s      │     38%            │
├──────────────┴──────────────┴───────────────────┤
│  Top Categories          │  Guardrail Triggers   │
│  1. Benefits (32%)       │  Salary block: 12     │
│  2. Leave/PTO (28%)      │  Legal redirect: 8    │
│  3. Remote Work (15%)    │  PII detected: 3      │
│  4. Organization (12%)   │  Confidentiality: 1   │
│  5. Compliance (8%)      │                       │
│  6. Other (5%)           │                       │
├──────────────────────────┴───────────────────────┤
│  Unanswered Queries (low confidence)             │
│  - "What's the policy on bringing pets to...?"   │
│  - "Can I switch from PPO to HDHP mid-year?"     │
│  - "Is there a sabbatical program?"              │
└──────────────────────────────────────────────────┘

Compliance Audit Log

Every interaction is logged for regulatory compliance:

{
  "timestamp": "2025-11-15T10:30:00Z",
  "employee_id": "EMP-008",
  "query": "What's our parental leave policy?",
  "category": "leave",
  "sensitivity": "normal",
  "guardrails_triggered": [],
  "model_used": "sonnet",
  "sources_cited": ["HB-001", "POL-001"],
  "confidence": "high",
  "latency_ms": 1200,
  "tokens_used": 850
}

This log serves multiple purposes:

Alerting Rules

Weekly HR AI Report

Automated weekly email to the HR leadership team:

This report is how HR gets value from the system beyond just answering questions — it reveals what employees are confused about, which policies need clarification, and where the knowledge base has gaps.

What You'll Build

Glossary