Tools Dashboard — Multi-Tenant SaaS Admin Platform

The Problem

The team was building several internal products, each of which needed its own auth, its own user model, and its own admin surface — which meant users logging in five times and engineers rebuilding the same plumbing five times.

The required surface was broad: a public-facing app, an admin panel, an auth service, a WebSocket service for real-time updates, background workers, and a registry that tracked which features belonged to which service. Traditional monolithic architecture couldn't meet the scaling requirement; naive microservices couldn't meet the "don't rebuild auth five times" requirement.

The Approach

A Kong 3.4 API gateway at the front door with JWT validation, a Web Application Firewall (SQL-injection, XSS, path-traversal rules), and tiered rate limits (60 / 600 / 10,000 req/min by plan). Behind it, FastAPI services: one for the public API, one for auth, one for WebSockets, one for the feature registry, and Celery workers for background jobs.

Auth combined Google OAuth, email/password with bcrypt cost 12 and a HaveIBeenPwned breach check, and TOTP 2FA enforced for admins — with single-use refresh rotation and CSRF tokens on every state-changing call.

Four data stores, each with a job: PostgreSQL for accounts and structured data, Cassandra for time-series events and an audit trail partitioned by user/tenant, Redis for cache and pub/sub fan-out, and SeaweedFS for S3-compatible object storage. The WebSocket service was engineered for 10,000 connections per server with a 10-msg/sec/user cap.

The standout piece was the App Library: users browse registered apps as cards, click one, and get auto-authenticated into the remote app via an OAuth redirect — making Tools Dashboard a mini identity provider for the whole product family.

The Outcome

A production-ready multi-service platform with unified auth, real-time updates across 14 services, and a single sign-on launcher for every app in the family. Every service shipped a per-service 'CONTEXT.md' and a feature-contract file, which is the reason a 14-service platform stayed navigable instead of collapsing into a sprawl.

Engineered against hard targets: GET under 100ms cached / 500ms uncached, POST under 1s, file upload under 5s, WebSocket up to 10,000 connections per node. WCAG 2.1 AA on the front end.