A production-grade, fully branded education platform, live payments, in-browser creative tools, microservices architecture, and zero lost enrolments. Here's exactly how we built it. For what the platform does for learners and teams, see the education platform page.
The client needed a platform no SaaS product could offer, complete ownership, custom branding, and a payment system they could trust with real revenue.
The Challenge
A growing education company had outgrown every generic LMS on the market. Teachable, Thinkific, Moodle, none of them could support a vision where teachers record professional lessons, build interactive mind maps, publish content to social media, conduct live whiteboard sessions, and sell courses and tool subscriptions, all under one roof, fully owned, fully branded.
They came to us with one clear requirement: build something that scales, doesn't break when payments happen, and is secure enough for real money. No compromises. No vendor lock-in. No half-measures.
What We Built
A production-grade, multi-service education platform serving teachers, students, and administrators with live payments, automated lifecycle emails, fuzzy course search, personalised recommendations, and a full in-browser creative suite.
System Architecture
Every service owns its domain, its database, and its deployment lifecycle. Cross-service communication is fully asynchronous via Apache ActiveMQ, a slow recommendation engine can never cause a payment to fail.
Full system architecture, React SPA · Spring Cloud Gateway · 7 microservices · Apache ActiveMQ · pgvector · AWS S3 / SES
Microservices
| Service | Responsibility |
|---|---|
| API Gateway | JWT validation, rate limiting (20 RPS / 40 burst), CORS, routing to all downstream services |
| User Service | Registration, login, JWT minting, password reset, email verification |
| Course Service | Course CRUD, chapters / modules / resources, ratings, pg_trgm fuzzy search |
| Learning Platform | Enrolment, per-lesson progress tracking, completion signals via ActiveMQ |
| Commerce Service | Cart, orders, Razorpay payments, tool subscriptions, coupons, HMAC verify |
| Media Service | File upload to S3, transactional email via AWS SES, lifecycle notifications |
| Recommendation Service | Personalised course recommendations, "Continue Learning" feed, vector similarity |
Engineering Depth
This isn't a naming convention, it's a hard boundary between business logic and infrastructure that paid off when we swapped the mock payment gateway for live Razorpay: zero domain code changed. One adapter file.
Enforce all business rules. ToolSubscription is a state machine, PAYMENT_PENDING → ACTIVE → EXPIRED / CANCELLED / PAYMENT_FAILED. No invalid state is representable in code.
Defined in the application layer, implemented in infrastructure. The domain has zero dependency on JPA, Spring, or any framework. Swap PostgreSQL for MongoDB, domain stays untouched.
Orchestrate use cases. They never reach into persistence or HTTP directly. BookAppointmentUseCase, EnrolStudentUseCase, each owns one workflow, nothing more.
Handle JPA entities, S3 uploads, JMS publishing, and Razorpay calls, fully replaceable without touching business logic. Hexagonal architecture enforced from day one.
Payments & Reliability
The most important architectural decision in the entire project. When a payment succeeds, two things must happen atomically: save the payment and trigger enrolment. If the message broker goes down between those two steps, the payment is recorded but the student is never enrolled. We solved this completely.
Razorpay confirms. Commerce service writes to outbox atomically with payment record.
Scheduled poller reads undelivered events from the outbox table.
Event forwarded to ActiveMQ. Marked delivered only on successful publish.
Learning platform consumes event and auto-enrolls the student.
If the broker goes down, the outbox accumulates. When it comes back, every event is delivered in order, exactly as written. Zero ghost enrolments since launch.
Every event listener is idempotent. A student receiving the same CoursePurchasedEvent twice is enrolled exactly once, no duplicates, no side effects.
Payments
Card data never touches our servers. Every payment callback is cryptographically verified server-side before any access is granted. Browser manipulation cannot bypass subscription checks.
POST /checkout, Commerce Service creates an Order + Payment record, calls Razorpay Orders API, returns razorpayKeyId + gatewayOrderId.
Razorpay Checkout modal opens in the browser. No card data ever touches our servers, PCI scope eliminated entirely.
Frontend calls POST /payments/verify. Backend computes HMAC-SHA256(keySecret, orderId|paymentId).
If signature matches, payment marked confirmed and outbox event fires. If it doesn't match, access is denied, regardless of what the frontend claims.
AI Recommendations
The recommendation service uses a hybrid approach, vector cosine similarity via pgvector as the primary engine, with pg_trgm similarity as a fallback for keyword and tag matching. Both layers combine with full-text search and business filters.
cosine_ops via pgvector
Results
Every architectural decision was made to protect revenue, reduce operational overhead, and deliver a platform the client's team can own and extend independently.
Draft → Teacher submission → Admin approval → Published, with rejection reasons and resubmission flow built in.
Server-side access checks on every tool page. Browser manipulation cannot bypass a paywall, only a valid DB row grants access.
Expired subscriptions at 01:00 UTC, reminder emails at 02:00, orphaned payments at 03:00, outbox cleanup at 03:30. Zero manual intervention.
Technology Stack
How We Build
Traditional recommendation systems rely heavily on tags and keyword overlap, this often misses deeper contextual relationships between learning content. Here's how we thought differently.
Tag and keyword-based matching surfaces courses that share labels, not courses that share meaning. A student interested in "neural networks" gets courses tagged "AI" but not the foundational "linear algebra" course they actually need next.
Semantic embeddings via Sentence Transformers + pgvector in PostgreSQL with cosine similarity search. Embeddings capture contextual meaning, not just surface keywords. Combined with event-driven ingestion pipelines for continuous freshness.
We optimised for operational simplicity, scalability, and cost efficiency without prematurely introducing unnecessary infrastructure complexity. pgvector with HNSW index delivers sub-100ms similarity search at our scale, no separate vector database needed.
The platform delivers more relevant recommendations, contextual learning discovery, and improved personalisation, through a scalable pipeline that evolves responsibly with scale. Good architecture isn't about using the most technologies. It's about building systems that grow cleanly.
Architecture Snapshot
cosine_ops
Good architecture is not about using the most technologies. It's about building systems that evolve responsibly with scale.
How We Build
As the platform evolved with browser-based recording, clip editing, timeline stitching and async exports, traditional in-memory processing using ConcurrentHashMap became a hard limit on horizontal scaling. Here's how we redesigned the architecture around Redis-based distributed coordination.
In-memory job state breaks the moment you run more than one server instance. Two workers can simultaneously pick the same job, a server restart loses all in-flight processing, and a client on instance A can't see the state of a job running on instance B.
Distributed coordination via Redis: LPUSH/BRPOP for job queues, atomic SET NX EX for distributed locking, HSET/HGETALL for shared job state, and Pub/Sub channels for real-time progress to clients via WebSocket.
Redis gave us sub-millisecond performance, atomic operations, TTL-based auto-cleanup, and Pub/Sub, all in a single lightweight infrastructure component. No Kafka or ActiveMQ overhead at this scale. Stateless workers mean we add instances with zero configuration change.
Multiple video worker instances run safely across servers without race conditions or duplicate processing. Clients receive live progress updates without polling. Lock TTL auto-expires on crash, jobs re-enter the queue automatically. Horizontally scalable from day one.
Architecture Snapshot
LPUSH / BRPOP video:jobsSET lock: {videoId} NX EX 30HSET job:{id} status progressINCR ratelimit:{userId} EX 60
Build lean initially. Design for scale from day one.
We build custom software for education companies, SaaS startups, and businesses that need more than a template can offer. Tell us what you have in mind, or see our full range of services.
Built for your exact workflow, not a generic LMS with your logo on it.
Tamper-proof, transactionally safe, and verified server-side. No lost transactions.
Designed for independent deployment, owned by your team, built to scale.
Recording studio, workflow, 3D, whiteboard and publisher, integrated into your product.