Apache-2.0 · Offline-first · Multilingual · Built on the public education commons

Give India's 1 crore teachers their hours back.

OpenPathshala is an open-source AI toolkit that automates the highest-friction classroom workflows — grading handwritten answers, setting question papers, generating mother-tongue notes — designed to run on every tier of Indian school infrastructure, from no-electricity rural schools to fully-connected urban ones, at a cost of single-digit rupees per task.

Live demo runs on open LLMs (Groq). The same router targets Claude in production.

14.71 L
schools
UDISE+ 2024-25
24.8 Cr
students enrolled
1.01 Cr
teachers
36.5%
schools with no internet

The central design constraint

One country, three completely different schools.

A national tool that assumes connectivity excludes the schools that need help most. Here's the real infrastructure picture from UDISE+ 2024-25 — the gap between "has electricity" and "has working internet" is the whole problem.

Source: UDISE+ 2024-25, Ministry of Education, Govt. of India. Functional %, national average.

Tiered degradation, not feature gating

The same workflow runs everywhere — it just degrades gracefully.

Premium schools don't get features poor schools are denied. Every tier gets the workflow; the model quality and latency adapt to what the school has.

T0

No power · No internet

Paper + a phone that's briefly charged. Capture now, queue, sync and process later. Fully on-device / edge open models — zero recurring API cost.

  • Photograph scripts when a device is free
  • Batch-grade overnight, print & return
  • On-device OCR + small open LLM
T1

Shared devices · Intermittent net

The majority of Indian schools. A single mini-PC at a Block Resource Centre serves a cluster. Open models do the bulk; a frontier model is reserved for hard judgment.

  • Edge node syncs when online
  • Open-model transcription + selective escalation
  • Per-school budget caps
T2

Connected · 1:1 or lab

Funded / private schools. Full cloud routing — the value (teacher-hours returned) dwarfs the cost. Frontier vision + reasoning on every task.

  • Real-time grading & feedback
  • Misconception analytics for the class
  • DIKSHA / WhatsApp integrations

See it in action

How AI does the heavy lifting.

Animated walkthroughs of the core workflows — from a phone photo of a handwritten script to a graded result, from a textbook to a ready exam paper, and a camera that watches an exam hall so a teacher doesn't have to.

Handwritten answer script AI reads it (HTR) script-id: 0427 · lang: Devanagari confidence: 0.93 vs marking scheme reactivity oxidation exothermic 8/10 graded draft
Snap a photo on any phone Instant graded feedback 4/5 confidence 91% ✓ method correct ✓ units shown ✗ final step missed मातृभाषा में फ़ीडबैक अच्छा प्रयास! अगली बार अंतिम चरण ज़रूर लिखें।
Textbook (online or scanned) Chapter 1 Blueprint constraints MCQ Short Long Σ = 20 marks · Bloom-spread Ready paper + key QUESTION PAPER 20 marks ✓
Live camera · exam hall ● REC ! AI proctor 24 present 22 on-task ⚠ seat C3 looking at neighbour Privacy: faces tokenised on-device · no cloud upload teacher confirms every flag
Photo of paper register # Name P/A 1 2 3 4 5 Structured digital roster Aarav S. Diya P. Kabir M. Meera J. Present 38 / 40
After-hours doubt — grounded strictly in the textbook Why is the sky blue, didi? — student, Class 7 thinking… AI tutor Sunlight scatters off tiny air particles. Blue light scatters most — so the sky looks blue. (Class 7 Science, Ch. 11) 📗 Grounded source no open-web guessing
A1 Photograph a stack of scripts, AI transcribes the handwriting, grades each answer against the marking scheme, and flags low-confidence ones for the teacher.

Live prototype — really runs

Try the two MVP workflows.

These call open LLMs server-side through the same 3-tier router the architecture specifies. Each result shows you which model tier handled it, the latency, and the per-task cost. Teacher-in-the-loop: every output is a reviewable draft, never a final mark.

Results appear here →
A blueprint-compliant paper appears here →
A simplified, mother-tongue explainer appears here →

⚠️ Shared public demo — rate-limited and capped. Outputs are AI-generated drafts for demonstration; in a real deployment a teacher reviews and overrides every result.

§7 of the spec — the most important cost decision

Route by judgment required, not by default to the biggest model.

Most classroom tasks are cheap extraction. A minority need reasoning. A small fraction need deep reasoning. Honest routing — a 70/20/10 split — cuts cost by more than half with negligible quality loss.

Prompt cachingup to −90% on the rubric/blueprint reused across a whole class set
Batch processing−50% — grading & overnight paper-gen aren't latency-sensitive
Open-model floorevery step has an open-weights fallback — the toolkit never hard-locks to one vendor

A toolkit, not a single app

The workflow catalogue.

Each module is chosen because it returns hours to a teacher — not because it's technically interesting. The two highlighted are the live MVP.

Built in the open

A genuine public good — Apache-2.0, built on the commons.

OpenPathshala extends India's public education infrastructure — DIKSHA / Sunbird, Bhashini, openly-licensed NCERT content — instead of rebuilding it. The toolkit must run with no paid API at all (open models are the floor); frontier models like Claude are the quality ceiling it routes to when a school can fund it.

  • ✅ Permissive Apache-2.0 license — maximizes adoption & dependency
  • ✅ Provider-agnostic router — Groq / Ollama / vLLM / Claude
  • ✅ DPDP-Act-aligned privacy: on-device redaction, consent ledger, teacher-in-the-loop
  • ✅ Offline-first so it reaches the schools that need it most

The path to credibility

Ship publicly → deploy to 5–10 real schools → measure teacher-hours saved & grading accuracy → apply via the discretionary "ecosystem depends on it" path with evidence.

P0 Grading engine + router
P1 Question-paper generation ← MVP live now
P2 Redaction + offline edge deploy
P3 Multi-board + analytics
P4 DIKSHA / UDISE / WhatsApp