AI Copilot

Overview

The AI Copilot is a production feature embedded in a multi-tenant SaaS, giving each organization a grounded assistant over its own data. It runs across the web application and WhatsApp, sharing one backend and one agent runtime.

Architecture

• Router + specialist agents — a lightweight router classifies intent and delegates to specialist agents, each exposing a focused set of tools via function calling.
• Grounded retrieval — a RAG pipeline over pgvector injects per-tenant context (documents, records, settings) with strict row-level isolation, so answers never leak across organizations.
• Provider-agnostic models — GPT-4o, Claude and Groq/Llama sit behind a single interface; the router picks a model by cost, latency and capability, with automatic fallback.
• Streaming everywhere — responses stream token-by-token to the web UI and are chunked for WhatsApp delivery, keeping perceived latency low.

Engineering focus

The hard parts were reliability and safety, not the demo: tool schemas that the model calls correctly, tenant isolation in retrieval, deterministic fallbacks when a provider degrades, and full observability (traces, costs, error tracking) so the feature could be trusted in production.