Embedded Full-Stack Pod for an EdTech SaaS Canvas Editor

Yes — embedded delivery is the default operating model. The Gradoo pod joined the client’s GitHub, Jira, sprint planning, code review, and release rituals from week one, with a Tech Lead acting as the primary client contact to absorb coordination overhead. Feature delivery continued through the engagement; weekly releases shipped throughout the canvas performance, export reliability, and collaborative editing work. Freezing product to rework a live SaaS canvas editor would not have been viable through the May–June submission spike — the phased, in-flight approach is what kept the platform serving 3,500+ school classes without service degradation.

Two engineering moves. Virtualised viewport rendering — only the elements actually visible (plus a small buffer) are mounted in the DOM, so a page with 20–50 images stops paying the render cost for every element on every interaction. And canvas state separated from React’s render cycle — drag, drop, and transform interactions update an isolated state store and reconcile to React on a controlled boundary, removing the cascading re-renders that produced the ~800ms lag. The combined effect brought interaction latency from ~800ms to under 80ms on dense pages without changing the canvas feature surface.

The original synchronous export ran inside the request lifecycle and timed out under spike load. Uvik Software decoupled PDF export into an async job queue with priority lanes, retry logic, and dead-letter handling for unrecoverable failures. The export now runs as a background workload sized to the spike profile rather than the steady-state load, with priority lanes so paying users on tight submission deadlines do not queue behind bulk jobs. Export success rate at peak moved from 78% to 99%+, sustained through the May–June submission window and the seasons that followed.

Optimistic locking with session-level element claiming, combined with GraphQL subscriptions for real-time state propagation. When a user selects a layout element to edit, that element is claimed for the session; concurrent attempts by other team members are blocked at the claim layer rather than colliding at write time. State changes propagate to all connected clients through GraphQL subscriptions so the UI reflects the locked state immediately. Concurrent-edit conflicts dropped to zero across the engagement window, and the change was invisible to users on the happy path — they only see the difference when a teammate is already working on the same element.

Layout elements (images, text blocks, decorative shapes, embedded media) share some properties and differ in others. A single-type schema either over-generalises (losing type safety) or proliferates with optional fields (losing clarity). GraphQL union types let each element type carry its own typed shape while still being queried through the layout’s element list. Aligned to MongoDB documents stored under the same union pattern, the schema becomes the source of truth for both API and storage. Adding a new element type — a previously 3-day task involving schema, resolver, validation, and storage changes — now ships in under 4 hours.

Most pods run 12+ months because the embedded model produces compounding velocity — the longer the pod stays inside the client’s product organisation, the more context it carries, the faster it ships. The Gradoo pod scaled with the client as class count tripled and the architecture held up cleanly under peak load. Exit paths are documented: if internal hiring catches up to the roadmap, Uvik Software unwinds the pod gradually with full handover documentation. No vendor lock-in; the client owns the architecture and the codebase end-to-end.