Scan. Understand. Group. Act. Reassess. Kagaz AI reads handwritten student work, maps it to foundational competencies, finds evidence-based learning gaps, and tells you what to teach next — with locally available materials.
In classrooms of 40+ students, grading manually leaves no time for conceptual diagnosis. Basic slips in place-value decimal carries or unit borrowings are masked behind quick red ticks.
Math Problem: 52 - 18 = ?
> Student Answer: 44 (Wrong. Red pen tick.)
The teacher marks the question as wrong, but has zero time to notice the student subtracted units from bottom-to-top (8 - 2) due to borrowing confusion. Gaps persist.
Math Problem: 52 - 18 = ?
> OCR Extracted Answer: 44
> Observed: 8−2 pattern suggests borrow error — needs verification
> Kagaz asks the teacher one quick verification question first
AI extracts the digits and maps them to competencies. One wrong answer is never called a gap — recurring patterns across items are, and the teacher confirms every conclusion.
Vision-language OCR (local Ollama with Groq cloud fallback) reads the handwriting; a deterministic, framework-driven evidence engine — not an LLM — decides what each child can do.
Canvas downscales on-device and streams image bytes directly into the backend memory pool to prevent I/O disk bottlenecks.
Constructs a strict multimodality prompt bounding the model to extract math geometry without hallucinating external context.
A local Ollama vision model (Qwen2.5-VL) reads the handwriting; if unavailable, Groq's cloud vision API takes over. No single point of failure.
Output logic mandates a deterministic JSON schema containing student answers, true answers, and diagnosis reasoning.
Response strings are sanitized to remove markdown code blocks and trailing delimiters before passing to the state manager.
Results are held in a draft state on the client. Teachers review the AI's diagnosis alongside the original image crop before committing.
Structured workflow designed for real classrooms
1. Capture Scans
Snap photos of student notebook pages or scan batch sheets.
2. Digitize Work
Our robust OCR extraction system transcribes handwritten student responses.
3. Understand Gaps
The evidence engine maps errors to competencies; recurring patterns get confidence labels, single errors get verification questions.
4. Remediate Work
Deploy targeted practice sheets and group plans automatically.
Real-time mapping of crop coordinates to database keys
Handwritten Paper Segment
Q1. Solve: 14 + 19
Student Answer: 23
Structured JSON Payload
{
"student_answer": "23",
"expected_answer": "33",
"is_correct": false,
"observation": "23 = 14 + 9: the ten from (4 + 9) was not carried",
"competency": "addition_no_regroup",
"confidence": "needs_verification",
"verify_by_asking": "What is 4 + 9?"
}Why teachers cannot simply use general-purpose vision tools like Google Lens or ChatGPT.
| Feature | Google Lens + ChatGPT | Kagaz AI Platform |
|---|---|---|
| Multi-Column Layouts | Reads left-to-right across columns, merging unrelated questions. | Auto-detects column centroids and segments questions cleanly. |
| Indic Pencil Handwritings | Fails on faint pencil marks and regional (Hindi/Devanagari) digits. | Vision models + teacher correction of every extracted answer. |
| Pedagogical Guardrails | Hallucinates diagnostic terms; suggests generic internet lessons. | Deterministic evidence engine maps results to an FLN/NCERT-aligned competency framework; every inference shows its evidence and confidence. |
| Classroom Scale & Connectivity | Uploads heavy multi-MB photos; fails instantly when offline. | Compresses images to 300KB on device; queues offline in IndexedDB. |
Interactive preview of workspaces built for teachers
Upload worksheet images. When horizontal scale is required or Redis is configured, tasks are offloaded to asynchronous Celery queues automatically to maintain server uptime.
Drop files here to start upload
Accepts JPG, PNG up to 5MB
How Kagaz AI transitions from a hackathon prototype to a highly scalable product
Our client-side compression keeps uploads small (a 5MB photo shrinks to ~300KB) so scans work on slow rural connections. The pedagogical pipeline runs on plain application code — no expensive per-image inference is needed for diagnosis.
Designed to integrate directly with public school databases (UDISE+) and partner with educational NGOs (like Pratham or Central Square Foundation) to roll out diagnostic tracking via existing CSR initiatives.
Free tiers for individual teachers in government schools, funded by premium institutional dashboard licensing for district administrations, school networks, and student diagnostic audit reporting.
Engineered to work reliably across rural government schools
Scan a full set of papers in minutes and get group-level insight the same class period — instead of hours of manual checking. (Time actually saved depends on your workflow; we measure it in-app.)
Queue captured sheets offline via IndexedDB. Your roster syncs automatically once an active connection is detected.
Advanced OCR models adapt to real student pencil writing, local numerical scripts, and standard notebook lighting.
Switch instantly between Hindi and English languages to accommodate varying teacher preferences and regions.
Built for real classrooms: scan student work, see evidence-based learning gaps, get the next teaching action.