[● PRODUCTION]
Built at Quantum Strides · sole AI engineer
Multi-Tenant White-Label AI Chatbot
A serverless, white-label chatbot platform with strict per-tenant isolation, self-serve onboarding, and an embeddable widget.
// overview
A multi-tenant SaaS chatbot any customer can drop onto their site. A WebSocket pipeline routes each message to that tenant's private knowledge base, synthesizes a grounded answer with Claude, and streams it back through an embeddable widget — with data isolation enforced at the vector and context layers.
// the problem
Customers wanted a branded support bot grounded in their own docs, without leaking data across tenants or doing any engineering. It had to onboard in minutes, isolate every tenant, and scale to spiky traffic without a server to babysit.
// architecture
01 · widget
Embeddable JS
via CloudFront
02 · socket
API GW WebSocket
router Lambda
03 · retrieve
Per-tenant KB
Qdrant namespace
04 · synth
Claude
via Bedrock
05 · state
DynamoDB
conversation log
// side rail
Tenant isolation
namespace + context
↳ PII redaction middleware + per-tenant rate limiting at the edge
// what i built
- ✓Serverless pipeline: API Gateway WebSocket → router Lambda → per-tenant Qdrant knowledge base → Claude synthesis, state in DynamoDB.
- ✓Embeddable JS widget served via CloudFront, with configurable connectors (REST, webhooks, Confluence, Notion) for tenant data.
- ✓Fully self-serve onboarding: admin portal, connector setup, and usage analytics — no engineering required per tenant.
- ✓Tenant data isolation at the vector-namespace and LLM-context layers, with query-time access controls.
- ✓PII redaction middleware before context injection and per-tenant rate limiting at API Gateway.
- ✓Zero-downtime blue/green deployments via AWS CDK.
// impact
0
downtime deploys
100%
tenant isolation
4
data connectors
// tech stack
PythonAWS LambdaAPI Gateway WSAmazon BedrockQdrantDynamoDBCloudFrontAWS CDKECS Fargate
next project
Streaming Multilingual Voice Agent