Milo — WhatsApp AI Agent with Human-in-the-Loop Guardrails

Product, Architecture & Full-Stack Engineering (End-to-End) · Private Fashion Brand Client · Alias Case Study

Client AI Product · 2026

Designed, built, and deployed a production WhatsApp and web AI assistant that answers team questions from a shared knowledge base, digests group chats, and enforces human approval before any mutation.

Channels (WhatsApp + Web)
2
Mutations staged & confirmed
100%
End-to-end architecture
0→1
Running in production
Live

The Problem

A growing fashion brand team repeatedly interrupted the owner on WhatsApp for routine operational answers, reminders, and schedule checks—while unverified AI answers or unreviewed writes to shared company knowledge risked costly mistakes.

Product Tradeoffs & Decisions

  • Required every state-changing action to be staged and explicitly confirmed in chat rather than allowing autonomous writes—eliminating silent AI side effects in production.
  • Gated facts extracted from group chats behind an owner approval queue before entering the company knowledge base, keeping the shared source of truth reliable.
  • Instructed the tool loop to search the knowledge base, check calendars, and escalate to the owner rather than hallucinating or guessing.

Technical Approach

  • Built a Claude API tool loop in Next.js, TypeScript, and PostgreSQL integrated directly with WhatsApp and a companion web interface, deployed with Docker.
  • Added configurable specialist agents with per-call token metering and a cost-visibility view in the admin dashboard.
  • Automated morning digests from team group chats and structured reminder/meeting workflows.

Cross-functional Execution

  • Scoped workflows directly from the team’s real WhatsApp communication habits, built the full stack end-to-end, and shipped into daily production.

Measured Outcomes

  • Live in production answering the client’s team across WhatsApp and web.
  • Established safe agentic patterns—staged confirmation, KB approval gates, escalation over guessing, and per-call cost accounting.

Technologies & capabilities

  • Next.js
  • TypeScript
  • Claude API
  • WhatsApp
  • PostgreSQL
  • Docker