amanrr48/prajna-v4-predictor

PRAJNA v4 — Exam Prediction Engine

Predictive Resource Allocation for JEE/NEET Aspirants.

Overview

PRAJNA v4 is a hierarchical exam prediction engine that predicts which topics will appear in upcoming NEET/JEE exams, using 48 years of historical exam data (23,119 questions).

Architecture

• 10-signal appearance probability model with hill-climbing optimized weights
• Parent gate: chapter-level prediction gates micro-topic scoring
• Subject-balanced reranking with exam-specific quotas
• 3-stage pipeline: Appearance × Weightage × Format

Signals

1. Recency-weighted frequency (exponential decay)
2. Appearance rate (fraction of years appeared)
3. Recent 3-year presence
4. Recent 5-year presence
5. Gap return probability (overdue topics)
6. Trend slope (10-year linear regression)
7. Cycle match (periodic reappearance)
8. Parent inheritance (chapter score passed to micro-topics)
9. Recency burst (dense recent appearances)
10. Dispersion (rewards consistent appearances)

Performance

• Chapter-level backtest (NEET, k=50): 94.4% precision, 75.0% coverage
• Micro-topic level (NEET, k=100): Combined score 0.63
• Backtest accuracy: 91% (averaged across 2019-2023)

Data

• 23,119 questions from 292 papers (1978-2026)
• 48 years of NEET + JEE Main + JEE Advanced
• 755 unique micro-topics across 143 chapters

Usage

from predictor_v3 import predict_chapters_v3, predict_microtopics_v3

# Chapter-level predictions
chapters = predict_chapters_v3("exam.db", target_year=2026, exam="neet", top_k=50)

# Micro-topic predictions
micros = predict_microtopics_v3("exam.db", target_year=2026, exam="neet", top_k=200)

License

Apache 2.0