Connect a model to transcribe images.
You can review text and view results here. To read a new image, connect your OCR server.
HANDWRITTEN TO DIGITAL
A new chapter for every అక్షరం.
Turn Telugu handwriting into editable text. Review with care. Export with confidence.
01
Bring your documentA scan, a photo, a piece of history.
02
Transcribe & reviewYour original, always alongside.
03
Make it yoursClean text. Open formats.
Original document
SAMPLE
12 అక్టోబర్ 2026
ప్రియమైన మిత్రమా,
మన తెలుగు భాష ఎంతో మధురమైనది.
ప్రతి అక్షరంలో ఒక అందం ఉంది.
మన జ్ఞాపకాలను పదిలంగా ఉంచుదాం.
పాత లేఖలు, కథలు, అనుభవాలు
రాబోయే తరాలకు అందిద్దాం.
అక్షరాలకు కొత్త జీవితం ఇద్దాం.
ప్రతి అక్షరంలో ఒక అందం ఉంది.
మన జ్ఞాపకాలను పదిలంగా ఉంచుదాం.
పాత లేఖలు, కథలు, అనుభవాలు
రాబోయే తరాలకు అందిద్దాం.
అక్షరాలకు కొత్త జీవితం ఇద్దాం.
ఇట్లు,
మీ మిత్రుడు
మీ మిత్రుడు
ILLUSTRATIVE TYPESet SAMPLE
NEEDS REVIEW
◈
Sample text for exploring the editor. Connect a model to transcribe your own documents.
BUILT FOR TELUGU. DESIGNED FOR PEOPLE.Verbatim by default · Human reviewed by choice
YOUR COLLECTION
Document library
Reviewed text saved on this device. Images are never stored in the library.
MEASURE, DON'T GUESS
Evaluation lab
Compare a prediction against a verified transcript. Punctuation and digits count.
NFC grapheme CER uses Unicode grapheme segmentation. WER splits on whitespace. These are text-pair metrics, not a model benchmark. The Python toolkit supports corpus evaluation.
PROVENANCE FIRST
Dataset registry
Source discovery is separate from downloaded, verified training data.
Historical pilot audit: 24,000 word samples and 3,491 labeled pages were technically verified. The preserved word pilot used 1,000 / 500 / 500 subsets from pools of 7,997 / 8,003 / 8,000 words. The active capped study uses the frozen splits shown above. Page training awaits clarification of dataset terms. Writer identities and source labels still need human verification.
OPEN WORKFLOWS
Built to work with your stack.
A small REST API, structured exports and reproducible Python tools.
Connect your model
Run the included API locally. Model loading requires an authorized IndicOCR checkpoint and its official dependencies. The UI works without a model for manual review.
python -m uvicorn ocr.server:app \ --host 127.0.0.1 --port 8765
Use the local website for a loopback API. Hosted HTTPS pages may block local HTTP connections.
Predict a page
curl -X POST http://127.0.0.1:8765/api/predict \ -F "file=@page.png"
The response includes raw Markdown, ordered blocks, model identity and a mandatory review flag. Layout confidence is never presented as transcription confidence.
Download API contract ↓Reproducible toolkit
python -m ocr.cli audit --root data --output manifests/all.jsonl --source my-corpus --license-ref CC-BY-4.0 python -m ocr.cli split --manifest manifests/all.jsonl --output manifests python -m ocr.cli evaluate --input predictions.jsonl --output reports/metrics.json
Includes image verification, NFC labels, SHA-256, visual duplicate grouping, deterministic group splits and grapheme-level evaluation. The original recognizer passed forward, backward, adapter update, tiny-subset learning and save/reload checks. See Training & results for the pilot status and measured metrics.