Project Kit / Product preview
AksharaNet OCR Starter Kit
A planned guided kit based on my Sanskrit Devanagari OCR project, connecting synthetic training data, line recognition and a web interface.
Explore the original project
This kit is in preparation. Explore the existing project and its documented scope below.
The problem
Printed Devanagari combines connected headlines, conjuncts and stacked matras that make character-by-character segmentation difficult.
The approach
Walk through the existing segmentation-free CRNN and CTC pipeline, from synthetic text rendering to page segmentation, Unicode decoding and API integration.
Who it’s for
Python developers and machine-learning learners exploring OCR for printed Sanskrit and Devanagari.
How it works
- Prepare printed-text samples and inspect the line segmentation.
- Trace CRNN predictions through CTC decoding into Unicode.
- Connect inference to the FastAPI backend and Flutter interface.
Planned package
- Planned architecture and training walkthrough
- Synthetic data generation and font setup notes
- Line segmentation and inference integration guide
- Evaluation checklist covering character error rate and corpus limitations
Proposed contents; the final package will be listed at launch.
Example output / Illustrative
Follow a page through the pipeline.
INPUT A page of printed Sanskrit PIPELINE Page → segmented lines Lines → CRNN predictions CTC decoding → Unicode text REVIEW Compare decoded text with a transcript. Record character errors and failure cases.
A sample of the intended format, not a measured result or a screenshot of a released product.
Before you buy
Is this available now?
Not yet. This is a preview of a planned release. No payment is collected here.
Which tools will it support?
Planned for PyTorch, CRNN + CTC, FastAPI, Flutter. Tested versions and setup instructions will be published with the release. Tool subscriptions, hosting and API usage are separate.
What about licensing and updates?
Planned usage covers personal and commercial projects, without redistribution or resale of the original package. Final license, release version and update policy will be published before checkout opens.
Is it right for a complete beginner?
Requires Python and machine-learning familiarity. The portfolio reports 14% line-level character error rate on held-out synthetic data. This is a learning kit for printed text, not a promise of perfect OCR or handwriting recognition.