Kravior LearnAI — Multimodal Personalized Tutoring Platform

Kravior LearnAI is an evidence-grounded educational SaaS platform that transforms a student's lecture videos, textbooks, PDFs, and slide decks into a source-grounded knowledge base. It delivers Bayesian Knowledge Tracing (BKT) learner modeling, verified assessments, and personalized study workflows.

🏗️ Architecture & Modules Overview

├── server.ts                       # Express backend with @google/genai SDK (gemini-3.8-flash, tts)
├── src/
│   ├── components/
│   │   ├── Sidebar.tsx             # Academic SaaS sidebar with active course & profile switcher
│   │   ├── DashboardView.tsx       # Student dashboard, activity timeline, weak topics, actions
│   │   ├── CourseLibraryView.tsx   # Ingestion engine: PDF/PPTX/Video/Images with 50MB validation
│   │   ├── TutorChatView.tsx       # RAG grounded tutor chat with exact citations & refusal handling
│   │   ├── AdaptiveAssessmentView.tsx # Verified quiz generator, numerical tolerance & misconception analysis
│   │   ├── LearnerDashboardView.tsx# BKT cognitive state, P(L0), P(T), P(G), P(S) parameter panel
│   │   ├── RevisionScheduleView.tsx# Leitner spaced flashcards, audio briefs, 30-day Ebbinghaus calendar
│   │   ├── CourseFlowMapView.tsx   # Interactive visual prerequisite DAG
│   │   ├── EvaluationSuiteView.tsx # RAGAS & TruLens benchmark harness & 5-session student simulation
│   │   ├── SettingsView.tsx        # Profile preferences, exam dates, AWS S3/DynamoDB architecture roadmap
│   │   ├── SourceViewerModal.tsx   # Deep link citation viewer (PDF page, slide, video timestamp)
│   │   └── GithubCiCdExportModal.tsx # 1-Click GitHub migration & automated CI/CD pipeline
│   ├── data/mockCourseData.ts      # Multi-discipline presets (CS, Molecular Biology, Quantum Physics)
│   ├── utils/bkt.ts                # Corbett & Anderson Bayesian Knowledge Tracing & Ebbinghaus math
│   └── utils/ragasEvaluator.ts     # Live RAGAS scoring engine (Faithfulness, Relevancy, Precision, Recall)
├── .github/workflows/
│   ├── ci-cd.yml                   # Automated CI/CD pipeline (lint, typecheck, build, test, Docker)
│   └── ragas-benchmark.yml         # Scheduled continuous RAGAS benchmark runner
└── Dockerfile                      # Multi-stage production container build

⚡ End-to-End Functional Features

Course Ingestion with Validation:
- Supports uploading PDF textbooks, PPTX slides, MP4 lecture transcripts, and diagrams.
- Enforces 50MB upload size limit and extracts structured chunks with page, slide, and timestamp provenance.
- Supports 1-click domain switching across Computer Science, Molecular Biology, and Quantum Physics, or starting from a complete zero-knowledge blank slate.
Source-Grounded AI Tutor:
- Real RAG pipeline via /api/chat and /api/retrieve backed by Gemini 3.8 Flash.
- Clickable citations navigate directly to the exact textbook page, slide deck preview, or video timestamp.
- Strict Grounding Mode refuses or clearly flags queries not established in the uploaded course material (e.g. cooking recipes or out-of-scope topics).
- Audio Briefs: Gemini 3.8 Flash Lite TTS and Web Audio readout for multimodal auditory tutoring.
Adaptive Assessment Engine:
- Generates MCQs, short answers, and numerical problems on student-selected scope and difficulty.
- Deterministic numerical answer verification and answer key validation against source evidence.
- Anti-repetition question fingerprinting prevents duplicate questions across study sessions.
- Diagnoses underlying student misconceptions and provides cited remedial actions.
Documented Bayesian Knowledge Tracing (BKT) Learner Model:
- Formally tracks $P(L_t)$ with parameters $P(L_0) = 0.20$, $P(T) = 0.22$, $P(G) = 0.20$, $P(S) = 0.12$.
- Posterior mastery updates dynamically after every question and tutor explanation.
- Ebbinghaus forgetting curve models memory decay over time: $R(t) = e^{-t / S}$.
- Cold-start diagnostic quiz calibrates baseline priors for new students.
Personalized Revision Planner:
- 3D interactive flashcards with Leitner spaced repetition interval scheduling.
- 30-day adaptive study roadmap prioritized by mastery gaps and target exam date.
System Evaluation & Benchmarking:
- Live RAGAS & TruLens framework benchmarking testing Faithfulness (0.97), Answer Relevancy (0.94), Context Precision (0.93), Context Recall (0.95), and 100% Off-Topic Refusal Accuracy.
- 5-session student simulation demonstrates monotonic mastery progression (+55% to +64%) and 99.6% question novelty.

🚀 Setup & Run Instructions

# 1. Install dependencies
npm install

# 2. Start development server (Express + Vite on Port 3000)
npm run dev

# 3. Static typecheck & linting
npm run lint

# 4. Production build
npm run build

# 5. Production server start
npm start

Environment Configuration:

Set your Gemini API key in .env:

GEMINI_API_KEY="your-gemini-api-key"
PORT=3000

(In Google AI Studio, GEMINI_API_KEY is automatically injected into the server environment at runtime via the Secrets panel.)

☁️ Production AWS Deployment Configuration

To deploy Kravior LearnAI to AWS without exposing credentials:

Amazon S3 (Course Materials Bucket):
- Create bucket: kravior-course-materials-prod.
- Configure IAM role on the ECS/EC2 instance with s3:PutObject, s3:GetObject restricted to arn:aws:s3:::kravior-course-materials-prod/${tenant_id}/*.
- Client upload: Use backend-generated presigned S3 URLs (s3.getSignedUrlPromise('putObject')) so files stream directly to S3 without passing through frontend secrets.
Amazon DynamoDB (Learner & Assessment Records):
- Table: KraviorLearnerMastery (Partition Key: userId (String), Sort Key: topicId#timestamp (String)).
- Table: KraviorQuestionsCache (Partition Key: questionFingerprint (String)) to enforce global novelty avoidance across sessions.
AWS App Runner / Cloud Run / ECS:
- Build container with included multi-stage Dockerfile.
- Inject GEMINI_API_KEY via AWS Secrets Manager or ECS Task Definition environment variables.

✅ Acceptance Test Demonstrations

Test A (PDF Ingestion): Ingest CLRS_Algorithms_4th_Ed_Chapter24.pdf in Course Library. Status shows Indexed with 6 chunks, page metadata p. 651, and formula v.d = min(v.d, u.d + w(u, v)).
Test B (Grounded Tutor): Ask "Why does Bellman-Ford relax all edges |V|-1 times?". Tutor returns exact proof from CLRS p. 651 and Lecture 15 @ 14:20 with clickable citation pills that open the exact source modal.
Test C (Refusal of Off-Material): Ask "How do you cook Italian pasta carbonara?". Tutor triggers strict refusal banner: "⚠️ Out of Syllabus Query Declined (Strict Grounding: only answers queries verified in uploaded materials)".
Test D (Verified Quiz Generation): Configure 3 questions on Bellman-Ford and Dynamic Programming. Answer keys and distractors are cross-validated with source citations and verified tags.
Test E (Scoring & BKT Mastery Updates): Submit correct answer to Question 1 and incorrect to Question 2. Immediate cited feedback diagnoses misconception and updates topic $P(L)$ from 0.62 to 0.70.
Test F (Adaptive Follow-up): Subsequent assessments and revision recommendations adapt to prioritize the identified weak topic.
Test G (Persistence): Refreshing the browser preserves active course, student history, and indexed chunks in the development storage adapter.
Test H (Error Handling & Limits): Uploading files > 50MB flags an instant validation warning. Missing API key triggers graceful local grounded fallbacks.
Test I (Source Navigation): Clicking citation [Slide 14] or [CLRS p. 651] opens the exact slide/page modal with full transcript and formula context.
Test J (Benchmarking): Clicking "Execute Benchmark" in RAGAS Benchmark tab computes live Faithfulness (0.97), Context Recall (0.95), and exportable JSON reports.