In-Product AI Assistant

Context

A global HR and employer-of-record platform holds payroll, leave, and workforce data across many roles and countries. Users needed answers and actions in one place.

Problem

A useful in-product assistant has to do far more than chat. It must take destructive actions safely, respect every user's permissions, and never leak data across roles.

My role

Architected and built the assistant essentially solo, backend and frontend.

Approach

• Designed a provider-neutral abstraction so OpenAI, Anthropic, or a local mock swap behind one interface.
• Built a streaming orchestrator with a tool-calling loop and 58 typed tools (30 read, 28 action) spanning payslips, timesheets, salary raises, leave approvals, permission changes, and hire requests.
• Hardened it for production with role-scoped access middleware, per-bucket rate limiting, prompt-injection guardrails, action-token confirmation for destructive operations, and full audit logging.
• Shipped the complete React panel: launcher, streaming message list, conversation history, action-confirmation cards, and authenticated download widgets.

Tech stack