Ask questions in Hindi, English, or Hinglish. Get step-by-step NCERT answers with page citations — entirely on your device. No internet, no cloud, no fees.
Powered by Qwen 2.5 1.5B quantized to Q4_K_M · ~1.2 GB RAM · Works on ₹8,000 phones
<5ms
Out-of-scope refusal
1.2 GB
Peak RAM usage
0 bytes
Data sent to cloud
3 भाषा
Hindi · English · Hinglish
Live Demo
See how the 3-tier gated pipeline handles real NCERT questions.
Student: "khana digest kaise hota hai body me?"
Digestion (pachan) ka process ye hai:
(NCERT Class 10 Science, pp. 97-101)
Student: "dil me khoon kaise flow hota hai?"
Synonym expansion: dil → heart, cardiac | khoon → blood, circulatory
Double circulation: Deoxygenated blood → Right atrium → Right ventricle → Lungs → Oxygenated blood → Left atrium → Left ventricle → Body...
(NCERT Class 10 Science, pp. 106-108)
Student: "Newton ka third law samjhao"
Yeh topic aapke Class 10 Chapter 6 (Life Processes) syllabus me nahi hai.
Hint: Newton's Laws of Motion Chapter 9 me hain.
Architecture
Every query passes through deterministic gates before touching the LLM. Out-of-scope questions are refused in under 5ms with zero model inference.
Student selects their active chapter (e.g., Class 10 → Science → Life Processes). All retrieval is scoped to this chapter only — no cross-chapter leakage.
Normalize Hinglish queries via synonym expansion (80+ mappings). Run BM25 search over SQLite FTS5. If score is below calibrated TAU threshold → immediate refusal in <5ms, zero GPU/NPU cycles.
Pass top 1-2 retrieved chunks (max 800 tokens) to Qwen2.5-1.5B via llama.cpp. Model answers strictly from the passage with NCERT page citations. If passage lacks the answer → emits OUT_OF_CURRICULUM refusal.
Why Offline
Works in villages with zero connectivity, on flights, or when data packs run out. The model runs entirely on-device.
No student query ever leaves the device. No telemetry, no analytics, no cloud servers. Complete privacy for minors.
No API tokens, no subscriptions, no per-query charges. One download, unlimited usage. Like carrying a teacher in your pocket.
Ask in the language you think in. The synonym engine maps colloquial Hindi (pachan, khoon, saans) to academic NCERT terms automatically.
Fits in ~1.2 GB RAM. Designed for 4-6 GB budget Android phones — the devices most Indian students actually own.
Only answers from NCERT textbook content. No hallucinations, no out-of-syllabus tangents, no dangerous misinformation. Every answer cites page numbers.
Under the Hood
Qwen 2.5 1.5B
Base model
Q4_K_M GGUF
~1.1 GB quantized
llama.cpp
Inference engine
SQLite FTS5
BM25 retrieval
LoRA r=16
Fine-tuning
MLX / Unsloth
Training runtime
ChatML
Prompt format
Python + Gradio
Prototype UI
Zero dependencies
No cloud SDKs
Evaluation
200-query test suite across 3 categories. No LLM-as-judge — span-matching against passages and exact refusal matching.
100
Standard NCERT questions
50
Hinglish paraphrases
50
Out-of-scope (100% refusal target)
Shiksha-Edge is being built in the open. Star the repo, contribute content, or bring it to your school.