Education Management

Adaptive learning requires modeling both the content domain and the student. Architecture: 1) Knowledge Graph: Structure the subject as a directed graph. Nodes = concepts (e.g., 'Linear Equations', 'Quadratic Equations', 'Polynomials'). Edges = prerequisites (must understand linear before quadratic). Each concept has multiple content items (video, text, practice) at different difficulty levels. 2) Student Model: For each student, maintain a mastery score per concept (0-100%). Updated using Bayesian Knowledge Tracing (BKT) or Deep Knowledge Tracing (DKT). Features: correctness, response time, hints used, content type effectiveness. Also tracks: learning style (visual/text/practice), optimal study time, attention span. 3) Recommendation Engine: At each learning moment, recommend next content using: a) Knowledge graph prerequisites (don't recommend quadratic if linear mastery < 70%). b) Spaced repetition (concepts due for review based on forgetting curve — Leitner/SM-2 algorithm). c) Difficulty calibration — serve questions slightly above current mastery (zone of proximal development). d) Content type preference — if student learns better from video, prefer video. 4) Assessment: Diagnostic test at start (Computer Adaptive Testing — difficulty adjusts based on answers). Formative assessment after each content unit. Item Response Theory (IRT) calibrates question difficulty. 5) Feedback Loop: Every interaction updates the student model. Weekly model retraining with new interaction data. A/B test different content and paths — measure learning gain, not just engagement. Duolingo example: Uses spaced repetition + Bayesian model per word. Estimated 34 hours on Duolingo = 1 semester of university language class.

High-stakes online exams (JEE: 12 lakh students, NEET: 20 lakh) are among the most challenging distributed systems. Design: 1) Exam Delivery: Browser-based locked exam interface (Safe Exam Browser or custom). All question data pre-loaded at exam start (no network dependency during exam). Answers saved locally + synced to server every 30 seconds. If network drops, local answers preserved. At exam end, final sync with hash verification. 2) Scale Architecture: Not all 12 lakh start simultaneously — stagger by center time zone. But peak: 5 lakh concurrent. Read-heavy during exam (load questions), write-heavy at sync/submit. Partition by exam center — each center connects to nearest regional server. CDN for static content (images in questions). 3) Security: a) Question paper security: Encrypted at rest, decrypted only at exam start time using time-locked cryptography. Different paper sets prevent sharing answers. b) Exam integrity: Randomized question order per student. Webcam proctoring (AI flags: face not visible, multiple faces, eyes away from screen). Locked browser prevents other apps. c) Result security: Response encryption with student key. Server-side integrity checks (response hash must match). 4) Fault Tolerance: Power failure at center: UPS backup (mandatory), exam timer pauses, extra time granted. Server crash: failover to replica. Student's local answers synced when connectivity restored. Individual system crash: reserve machines at each center for swap. 5) Result Processing: After exam window (may span multiple days): collect all responses, apply answer key, normalize across sessions (percentile method — score = 100 × (students scoring below you / total students)). Handling disputes: if answer key challenged, recompute for all students. Publish results + detailed scorecard.

LMS data model centers on the relationships between students, courses, content, and assessments. Core Entities: 1) Course: {id, title, department, instructor, semester, status}. Contains Modules, each containing Content Items (video, PDF, quiz, assignment). 2) Enrollment: {studentId, courseId, role, status, enrollDate}. Tracks which students are in which courses. 3) Content Interaction: {studentId, contentId, startTime, endTime, completionPercent, score}. Every interaction logged — video watch time, pages read, quiz attempts. 4) Assignment: {courseId, title, dueDate, rubric, maxScore}. Submission: {assignmentId, studentId, submitTime, file, plagiarismScore, grade}. 5) Grade: {studentId, courseId, component, score, weight}. Final grade = weighted sum of components (assignments 30%, midterm 20%, final 50%). Key Metrics: a) Engagement: DAU/MAU (Daily/Monthly Active Users), average session duration, content completion rate. b) Learning Outcomes: Average score per assessment, pass rate, grade distribution (bell curve analysis). c) At-risk indicators: students with < 25% completion at mid-semester, declining login frequency, assignment non-submission. d) Content effectiveness: completion rate per content item, correlation between content engagement and exam scores, video drop-off points. e) Instructor metrics: average grading turnaround time, discussion response rate, content update frequency. Analytics Architecture: Every student action → event stream (Kafka) → data warehouse (BigQuery) → dashboard (Metabase). ML model predicts at-risk students based on early engagement patterns — alert instructor for intervention.

DIKSHA (Digital Infrastructure for Knowledge Sharing) is India's national education platform built on the open-source Sunbird framework. Architecture: 1) Sunbird Framework: Open-source building blocks — Sunbird Learning (content management), Sunbird Inquiry (assessments), Sunbird Obsrv (analytics), Sunbird RC (registry/credentials). Each is an independent microservice. Any state or institution can use Sunbird to build their own education platform. 2) Content Architecture: Content is tagged with: Board (CBSE, State Board), Grade (1-12), Subject, Topic, Medium (English, Hindi, 22 languages). Teachers from across India contribute content — reviewed by state quality teams. Content types: video, PDF, interactive HTML5 (H5P), assessments. Stored on CDN with offline support. 3) Scale: 30M+ daily active users during school hours. 5B+ cumulative learning sessions. Spikes during exam seasons. Architecture: content served via CDN (no origin hit for cached content), user data partitioned by state, analytics pipeline processes 500M+ events/day. 4) Telemetry: Every user action generates telemetry event: content played, page scrolled, question answered, search performed. Events streamed to analytics pipeline (Kafka → Spark → Druid). Dashboards show: state-wise adoption, content usage, learning outcomes. 5) Federated Model: Central platform with state-specific content and branding. Each state controls their content and user management. Single sign-on across states using DigiLocker/Aadhaar. 6) Why open-source matters: Sunbird adopted by Ethiopia, Brazil, and other countries for their education platforms. No vendor lock-in — government owns the platform. Community contributions accelerate development.

Video is the primary content format for EdTech — delivering high-quality video to millions of students across India's varied network conditions is a major technical challenge. Architecture: 1) Content Pipeline: Raw video (studio recording) → transcoding (multiple bitrates and resolutions): 240p (low bandwidth), 480p, 720p, 1080p. Formats: HLS (HTTP Live Streaming) for iOS, DASH for Android/web. Audio tracks: English, Hindi, regional languages. Subtitles in multiple languages. Total: one video → 20+ variants. Tools: AWS MediaConvert, FFmpeg. 2) Adaptive Bitrate Streaming (ABR): Player starts at medium quality. If bandwidth drops (student on 3G), automatically switches to lower bitrate — no buffering. If bandwidth improves (switched to WiFi), upgrades to HD. Client-side algorithm measures: download throughput, buffer length, and RTT. 3) CDN Strategy: Multi-CDN (CloudFront + Akamai + local CDN). CDN PoPs in 20+ Indian cities for low latency. Most popular content pre-cached at edge locations. Long-tail content fetched from origin on demand. Cost optimization: 80% of views on 20% of content (pre-cache the popular 20%). 4) Offline Download: Students download videos over WiFi for offline viewing (save mobile data). DRM protection (Widevine/FairPlay) prevents unauthorized sharing. Download manager with queue, pause/resume, quality selection. Storage management: auto-delete watched content after 30 days. 5) Analytics: Video engagement tracking: play, pause, seek, replay, speed change, completion. Drop-off analysis: if 40% of students drop at minute 7 of a 15-minute video, content team investigates — is the explanation confusing? Heat map shows which parts are rewatched (indicates difficult concepts). 6) Live Streaming: Unacademy/PhysicsWallah live classes: WebRTC/RTMP from teacher → media server → CDN → students. 100K+ concurrent viewers per popular class. Chat and Q&A alongside video. Recording for later playback.