Back to case studies
AI Workflow Infrastructure

Snitch: The Multi-Agent Support Swarm

Automated customer support architecture for India's leading D2C fashion brand.

How we built a collaborative ecosystem of 5 specialized AI agents sharing memory to resolve 94% of Snitch's e-commerce customer support issues automatically.

94%

Auto-resolution rate

47s

Avg resolution time

4.8/5

CSAT score

83%

Manual support workload cut

MemoryShared

Router Agent

Classifies messages and orchestrates the swarm

Product Expert

Handles sizing, materials, and product recommendations

Order Tracker

Connects to Shiprocket API for real-time delivery status

Billing Agent

Handles refunds, invoicing, and payment queries

Escalation Manager

Detects frustration and manages human handoffs

The Challenge

Support tickets do not follow a simple script.

Snitch, one of India's fastest-growing D2C men's fashion brands, was receiving over 800 support messages daily on WhatsApp. Their 6-person support team was permanently backlogged, causing response times to balloon to over 4 hours. Customers, frustrated by slow shipping updates and return delays, began canceling orders and leaving negative reviews.

Standard chatbots could not handle the complexity. A single support ticket often involved multiple intents: check shipping status, request a return code, complain about sizing, and demand a refund. Single-prompt AI bots would hallucinate, run out of system context, or fail to link up with Shiprocket and payment gateway systems.

The solution required an enterprise-grade multi-agent architecture: a team of specialized AI agents, each master of a single domain, working in concert to handle customer queries end-to-end.

Swarm simulation

Watch the agents communicate in real-time

See how a complex query involving shipping status, cancellation policies, and user sentiment is routed and resolved.

1. Intent Classification

The Router Agent extracts core actions, determines if APIs are required, and coordinates dependencies.

2. Collaborative Memory

Agents store context variables (such as verified order details) in a thread memory cache, keeping API calls fast.

3. Tone Control & Polish

The final response is formatted to feel natural and empathetic, removing all technical jargon or robot-speak.

Snitch Support Chat
hi, my order SN-48291 has not arrived yet. can i cancel it or get a refund?
Step 1 of 6
Orchestration Logs
Active
[SYSTEM] Session initiated via WhatsApp Webhook.
Thread: refund_swarm_session

Core Modules

Inside the support ecosystem

Dynamic Intent Router

First line of defense. Processes queries in 400ms, categorizes intents, and routes immediately to the specialized sub-agent with proper memory payloads.

Product Consultative AI

A consultative agent trained on materials, fit guidelines, and catalog data. Suggests items, matches sizes, and handles cross-selling naturally.

Courier API Integration

Direct connection with logistic partners. Tracks delays, updates dispatch locations, and generates shipping labels dynamically.

Invoicing & Refunds Handler

Cross-checks order IDs against transaction details. Prepares credit notes and processes refunds automatically through safe banking gateways.

Sentiment Escalation Manager

Scans for angry language, caps lock, or frustration tags. Automatically formats a human handoff ticket containing the full swarm chat summary.

Shared Memory Buffer

Allows seamless agent-to-agent transfers. If a query shifts from tracking to billing, the context is preserved so the customer never repeats details.

Architecture

Built for scalability

OrchestrationLangGraph Multi-Agent Workflows
AI / LLM ModelsGemini 1.5 Pro + Gemini 1.5 Flash
Vector DatabasePinecone (Semantic Product Catalog)
IntegrationsShiprocket API + Razorpay Webhooks
ChannelsWhatsApp Business API via Interakt
AnalyticsPosthog + Custom Dashboard

Results

Automation that drives sales

Within 14 days of deploying the Multi-Agent Support Swarm, Snitch saw ticket resolution rates soar. 94% of standard inquiries (tracking, sizing, billing) were resolved without human interaction. The average customer wait time crashed from 4 hours to just 47 seconds.

By automating the repetitive queue, Snitch was able to downscale their support staff from 6 active agents to a single manager overseeing the AI dashboard. CSAT scores increased to 4.8/5, and customer retention metrics improved by 18% in the first month.

94%

Resolution rate

47s

Resolution speed

4.8/5

CSAT score

83%

Savings in support costs

Drowning in support tickets?

Stop losing customers to slow reply times. Let us engineer a specialized support swarm that scales with your growth.