Skip to content
LEAPDocs
2026.08

Platform architecture

How the pieces fit: a React application, a domain-structured API, background workers, and managed AWS underneath.

The shape of the system

Three deployed services behind one load balancer, plus the managed data layer. The web application is a static bundle; the API and the worker are separate services so that a long-running generation job can never starve an interactive request.

Web application

React with TypeScript, built by Vite and served as a static bundle. Routing is client-side, split into four role-scoped route groups plus the public routes: the guest quiz taker, the phone capture handoff, and this documentation portal.

  • Server state is cached and revalidated per query key, so navigating back to a screen does not refetch everything blindly.
  • Proctoring runs entirely on-device in a web worker — face and voice signals become abstract flag events; frames and audio never leave the browser.
  • The bundle carries no institution-specific content. Branding is configuration, resolved at runtime.

API

An Express application organised by domain rather than by technical layer — each domain owns its router, service, repository and schema.

There are around forty domains, from course and assessment through to ivr and paper-upload. A request enters at a router, is validated against a schema, executes in a service, and touches the database only through that domain’s repository — which is where institution scoping is enforced.

Requests are rate-limited per user and per IP, and LLM-backed endpoints carry a much tighter limit of their own because their cost is measured in tokens rather than milliseconds.

Workers and jobs

Anything that takes longer than a request is a job. The queue is durable across deploys.

Generation work — syllabus extraction, lesson authoring, materials, grading, concept tagging — is submitted as a job and picked up by a worker service. Jobs resume from their last completed gate instead of restarting when a worker is replaced, so a deploy mid-generation costs minutes rather than the whole course.

Progress streams back to the browser, which is why the syllabus screen can show a live, staged view of what the pipeline is doing rather than a spinner.

Data layer

Aurora PostgreSQLThe system of record. Serverless with a warm floor, so cost tracks usage without a cold start on the first morning request.
RedisJob queue and short-lived coordination state.
S3Uploaded syllabi, scanned answer sheets, lesson recordings, generated slide decks and audio. Served through the application, never as a public bucket.
CognitoIdentity. The application holds a bearer token rather than a host-locked cookie, which is what lets a tenant subdomain and the apex share one session model.

Infrastructure

Everything is defined in Terraform — network, database, compute, storage, identity, WAF, observability and DNS are separate modules, so a change to the certificate policy is reviewable independently of a change to task sizing.

  • Containers run on ECS Fargate behind an Application Load Balancer, with per-service autoscaling.
  • A WAF sits in front with global rate rules; ALB access logs are retained.
  • GuardDuty and CloudTrail are enabled, with alarms routed to a notification topic.
  • A wildcard certificate covers per-institution subdomains, so adding a tenant does not require a certificate change.