Tutor Pro
“Built by a tutor, for tutors. A full-stack production system that automates every operational burden.”
Overview
Most tutors juggle spreadsheets, messaging apps, and manual bank transfers to run their classes. Tutor Pro replaces all of that with a single cohesive platform: automated scheduling, real-time notifications, live WebRTC teaching rooms, AI-generated feedback, and financial-grade invoicing — deployed and running in production. This is not a tutorial project. It solves real problems I faced as a part-time tutor, built to a standard I'd be comfortable deploying for other tutors to use.
Engineering Highlights
#1High-Performance Bulk Calendar Engine
Initializing an entire month of sessions for 30+ students one-by-one is unusable in practice.
Optimistic Batch Processing with In-Memory Deduplication (O(1) conflict detection), Single-Pass Database Query, and JDBC Batch Inserts for transactional integrity.
300+ sessions generated across a full student roster in < 800ms.
#2Event-Driven Real-Time Notifications
Students need to know instantly when a session is rescheduled, a grade is posted, or an exam is assigned — without polling.
Server-Sent Events over Spring Application Events. SseEmittersManager maintains a thread-safe ConcurrentHashMap of emitters per user, with a heartbeat protocol to survive Docker/proxy connection drops and auto-reconnect on the client.
End-to-end delivery in < 500ms. idx_recipient_read index keeps unread-count queries at < 10ms regardless of volume.
#3AI Feedback Engine (Groq + Llama 3.3 70B)
Writing individualized, contextually appropriate feedback in Vietnamese for every student after every class is time-consuming and repetitive.
Groq's inference API with prompt engineering that enforces Vietnamese tone, cultural context, and length constraints based on the student's performance data passed in at runtime.
Context-aware, culturally appropriate Vietnamese feedback in < 300ms — replacing static templates with genuine intelligence.
#4Financial-Grade VietQR Invoicing
Manual payment tracking creates reconciliation errors and time spent chasing parents for confirmation.
Dynamic VietQR generation embedded directly in PDF invoices, built to the NAPAS-247 standard. CRC-16 checksum validation ensures transaction data integrity end-to-end.
100% automated reconciliation flow. Error rate reduced from ~15% to near zero.
#5Interactive Live Teaching Room (WebRTC + WebSocket)
Teaching online requires more than video — synchronized whiteboard, session recording, and graceful hardware fallback.
Full-Mesh WebRTC for P2P media (server never touches the video stream). STOMP over WebSocket for state sync. Whiteboard throttled at 50ms cuts congestion 80%. Media Access Guard catches NotReadableError and surfaces human-readable guidance. Composite Stream Recording merges screen, webcam, and audio into a single .webm — entirely in-browser.
< 200ms interactive latency. One-click recording. No crash on hardware conflicts.
#6Drag-and-Drop Calendar with Prefetching
Calendar UIs feel slow when navigating months, and drag interactions drop frames on re-renders.
@dnd-kit with Optimistic Rollback. React.memo + useCallback eliminate 95% of unnecessary re-renders. Adjacent months prefetched automatically on load.
Month navigation instant (~0ms). Drag interactions at 60fps.
#7Intelligent Assessment & Auto-Grading
Digitizing exam questions from Word/PDF is tedious. Exam data loss from browser crashes or accidental navigation is unacceptable.
ExerciseParserService uses a Hybrid Regex Parsing Engine to bulk-ingest MCQ and Essay questions from plain text. Exam Player implements debounced auto-save (~500ms), timer sync, and auto-submit on expiry — fault-tolerant by design.
Question digitization time reduced by 90%. Zero data loss architecture.
Performance Benchmarks
Measured in production against real usage:
| Module | Metric | Before | After |
|---|---|---|---|
| Calendar | Initial load | ~1.8s | < 0.8s |
| Calendar | Month navigation | ~500ms | ~0ms |
| Finance | Page load | ~2.5s | < 0.8s |
| Finance | Data consistency | Desync issues | 100% accurate |
| Learning | Lesson detail query | 3 queries (N+1) | 1 query |
| Notifications | Delivery latency | — | < 500ms |
| Bulk Calendar | 300+ session init | — | < 800ms |
Architecture Decisions
Three choices that shaped every other decision in the codebase.
Modular Monolith over Microservices
Why: Solo project — microservices would mean 10× the DevOps overhead with no team to justify it. Clear module boundaries make extraction straightforward if scale ever demands it.
SSE over WebSocket for notifications
Why: Notifications flow in one direction. SSE is HTTP-native, proxy-transparent, and requires no handshake upgrade. WebSocket is reserved for the live room where true bidirectionality is actually needed.
React Query over Redux
Why: All async state is server state. React Query's stale-while-revalidate, optimistic updates, and cache invalidation handle this natively. Redux would be ceremonial overhead with no real benefit.
Security
Defense-in-depth across every layer:
JWT + Refresh Token via HttpOnly cookies
Dynamic RBAC — RoleEntity + Permission enum, @PreAuthorize per endpoint
Bucket4j on login and session creation endpoints
Zod (frontend) + Jakarta Validation (backend)
Roadmap
- Multi-tenancy — expand to support multiple tutors (SaaS model)
- Bank API integration — auto-confirm payments when funds arrive
- Zalo/Telegram bot for automated debt reminders
- Learning analytics dashboard with progress trend prediction