Artimus — AI-Powered Exam Platform

Overview

Artimus is an event-driven examination and analytics platform engineered for banking and competitive test preparation. It handles high-concurrency exam attempts, real-time analytics rollups, and asynchronous AI-assisted question generation.

To isolate write-heavy exam submissions from analytics and AI jobs, the system is architected as an event-driven microservices platform inside a Turborepo monorepo.

System Architecture

The system decouples responsibilities into dedicated Fastify microservices behind an API Gateway, communicating asynchronously over Kafka (Redpanda) and caching active session state in Redis.

Microservices Breakdown

Service	Primary Stack	Core Responsibility
gateway	Fastify, Redis	Reverse proxy, rate limiting, request validation, token denylist verification
identity-svc	Fastify, Prisma	User lifecycle, authentication, opaque token generation, blast-radius isolation
catalog-svc	Fastify, Gemini API	Question bank, subjects, chapters, and async AI question generation
assessment-svc	Fastify, Vitest	Timed attempt lifecycle, immutable question snapshots, authoritative scoring
analytics-svc	Fastify, MinIO	Kafka consumer read-models, percentile calculations, CSV export worker
notification-svc	Fastify, Web Push	Real-time announcements, browser push notifications, Server-Sent Events (SSE)

Key Technical Decisions

1. Schema-per-Service Database Isolation

Instead of provisioning five separate PostgreSQL database clusters, Artimus runs a single PostgreSQL 16 instance with dedicated schemas (identity, catalog, assessment, analytics, notification) and restricted database user roles. Each service maintains its own Prisma schema, migrations, and connection pool. Cross-schema joins are physically revoked at the database level.

2. Transactional Outbox Pattern

When an assessment attempt completes, an event must be published to Kafka without distributed drift. Outgoing events are written to an outbox table inside the same local database transaction. A dedicated worker drains the outbox to Kafka, ensuring guaranteed at-least-once delivery paired with idempotent Redis consumer handlers.

3. Authoritative Server-Side Scoring & Snapshots

When a student starts an exam, an immutable snapshot of questions and grading criteria is written to attempt_snapshot. Scoring is executed server-side in assessment-svc with pure deterministic functions covered by unit tests, making client-side answer tampering impossible.

4. Asynchronous AI & Export Workers

Heavy tasks like Google Gemini quiz generation and large CSV analytics exports are offloaded from request handlers. Endpoints immediately return job IDs while background workers process the tasks and write artifacts to MinIO object storage.

Monorepo & Infrastructure

• Monorepo: Turborepo with pnpm workspaces
• Shared Packages:
  • packages/contracts: Zod DTO and event schemas (single source of truth)
  • packages/kafka-kit: Transactional outbox publisher & idempotent consumer
  • packages/redis-kit: Redis client and centralized key management
  • packages/observability: Structured Pino logging and OpenTelemetry trace propagation
• Infrastructure: Docker Compose (docker-compose.yml), Caddy reverse proxy with automatic TLS