Zanira AI Customer Assistant

Problem

Zanira's customers reach out across WhatsApp, the website, and Instagram expecting instant, accurate answers on stock, sizing holds, order status, and payment (M-Pesa, cash-on-delivery). Answering manually across three channels doesn't scale, and generic chatbots aren't safe to trust with live inventory or payment details.

Solution

Built a LangGraph ReAct agent (create_react_agent) with 17 tools spanning product search/stock/colors, order status, customer/loyalty lookup, active offers, and WhatsApp-specific flows like size-hold creation/extension, COD finalization, and M-Pesa payment details. The agent connects directly to Zanira's Postgres database via psycopg2 connection pooling and raw SQL for low-latency reads, while all writes and business logic stay centralized in the Django backend: one source of truth, no data drift. Conversation memory runs on Redis, with an atomic Lua script for history append (safe under concurrent read/write) and a TTL-tracked "silence" key that triggers WhatsApp follow-up nudges after 4h/8h/12h of no reply. A rule-based "simple responder" short-circuits common queries before they ever reach the LLM, keeping cost and latency down. The whole thing runs as a FastAPI + Uvicorn service, Dockerized and deployed on Railway, alongside the Django REST backend, a Next.js 14 storefront, and a separate Next.js admin dashboard.

Outcome

Currently handling 31 customer conversations live across WhatsApp, website, and Instagram (29 active, 2 escalated to a human). That's a ~94% autonomous resolution rate across 28 unique customers, all routed through one unified inbox.

Engineering highlights

• Diagnosed and fixed a stale-context reliability bug: the model trusted its own earlier stock claims instead of re-querying inventory. A single upfront instruction wasn't enough, so it was fixed by repeating the re-check instruction alongside the live message.
• Added an allowlist guard on tool-generated image URLs before they reach WhatsApp, preventing a hallucinated or injected URL from being sent to a customer.
• Cost-aware routing: a rule-based responder handles common queries before any paid model call.
• Obtained WhatsApp Cloud API and Meta (Instagram messaging) developer access, worked from the provider docs, and tested webhooks and message flows against test numbers before going live.
• A parallel non-AI "Smart Accountant" module in the Django backend (fabric/labour/overhead costing, margin-targeted pricing suggestions, sales-velocity restock alerts, P&L/cash-flow reporting) shipped alongside the AI work, showing this isn't just LLM plumbing.