Vision & Motivation
Educational challenge: one-size-fits-all fails to engage; emotion and context matter
Our goal: empower each learner with a data-driven AI tutor that adapts in real time
Research basis:
- Picard's Affective Computing (1997) emotion as core to intelligence
- 2024 systematic review on AI-driven emotion assessment in education
Competitive edge: combining cognitive, affective & behavioral data
Technical Architecture
Clients
iOS (Swift), Android (Kotlin), Web (React)
API Gateway
HTTPS, JWT auth
Orchestration Service
Python/FastAPI
Ensemble Engines
- Rule-Based (Python/Flask)
- ML Model (Python + scikit-learn/PyTorch)
- LLM-Gen (Node.js + OpenAI)
Storage
PostgreSQL + JSONB; S3 for images
Data Pipeline & Analytics
Data Schema
id SERIAL PRIMARY KEY
user_id UUID
resume_data JSONB
sentiment_data JSONB
...
engagement_metrics JSONB
performance_dynamics JSONB
context_signals JSONB
...
sora_avatar_url TEXT
status TEXT
Asynchronous tasks (Celery/RQ) update JSONB fields
Key Metrics
- Avg. time-on-task, hint-use rate, scroll depth
- Mood scores (1–5)
- Completion rate vs. control
- Sentiment lift (Δhappy score)
- Dropout reduction
Adaptive Recommendation Logic
Feature engineering
Normalize all signals to 0–1
Clustering
K-means user-style segments
Rule-based policies
if (hint_rate > 0.5 && avg_time < 30) {
nextLesson.difficulty = 'easier';
}ML layer
XGBoost classifier/regressor
LLM integration
GPT-4 for context-aware explanations
Pilot Testing & Ethics
A/B Testing Plan
Test Groups: Group A (Core RAPID GROWTH) vs. Group B (RAPID GROWTH + telemetry signals)
Ethics & Privacy
- Opt-in photo analysis; encrypted buckets
- JSONB telemetry only, no raw logs
- Bias mitigation in sentiment models
- Cultural validation for affective cues
