Guidance and development estimate for Kazakh, Russian, and English OCR

[baytasinva-dotcom]

Hello RapidOCR team,

I represent BAYTAS-INVA in Kazakhstan. We are developing TULGA PRO READER, an assistive reading device for people with visual impairments. It runs on a Windows laptop with a document scanner and needs to work offline.

I found the PP-OCRv5 Cyrillic recognition model in the RapidOCR documentation, and PaddleOCR lists Kazakh, Russian, and English among the model's supported languages. Could you please advise:

1. What is the recommended RapidOCR configuration for recognizing Kazakh, Russian, and English printed documents, including Kazakh-specific Cyrillic letters?
2. Can accuracy be further improved for clean A4 scans by fine-tuning or custom language data? We can provide representative sample scans.
3. Does the RapidOCR team or a developer you recommend offer paid customization? If so, what would be the approximate cost and development/testing timeline?

We would appreciate any guidance on the model and setup to test first, as well as an estimate if custom work is needed.

Thank you.

[ub3dqy]

Hi, I tested RapidOCR 3.10.0 with the PP-OCRv5 Cyrillic mobile model on Windows CPU using local ONNX files. On generated A4 pages, it made four character substitutions in 214 Kazakh characters. The Russian and English samples matched the reference text. All nine Kazakh-specific letters are in the model's character set.

I can set up and benchmark offline OCR on up to 10 anonymized scans with reference text. The pilot includes Windows setup steps, error rates by language, and processing times. It would cost 120 USDC on Base and take two working days once we agree on scope and have the samples. The results can guide any fine-tuning.

Would that budget and payment method work for BAYTAS-INVA?

Dima Mak / ub3dqy

[baytasinva-dotcom]

Hi Dima, Thank you for your detailed response and testing. Could you please confirm whether PP-OCRv5 Cyrillic mobile in RapidOCR 3.10.0 is an already trained, ready-to-use model, and that your proposed 120 USDC pilot covers installation and testing of this existing model rather than additional model training or fine-tuning? We are interested in evaluating its offline accuracy for Kazakh, Russian and English as part of our TULGA reading machine for blind users. Please also explain how we can pay you the 120 USDC: - Which payment methods do you accept? Is bank transfer or another conventional method possible? - If payment must be in USDC on Base, please provide your payment details and step-by-step instructions. - Is payment made before or after the test, and can you issue an invoice or payment confirmation? Once clarified, we can discuss the sample scans and next steps. Best regards, Jomart Kaldybayev BAYTAS-INVA | Kazakhstan TULGA project

…

On Wed, 11 Nov 2026 14:57:32 -0800, Dima Mak ***@***.***> wrote: Hi, I tested RapidOCR 3.10.0 with the PP-OCRv5 Cyrillic mobile model on Windows CPU using local ONNX files. On generated A4 pages, it made four character substitutions in 214 Kazakh characters. The Russian and English samples matched the reference text. All nine Kazakh-specific letters are in the model's character set. I can set up and benchmark offline OCR on up to 10 anonymized scans with reference text. The pilot includes Windows setup steps, error rates by language, and processing times. It would cost 120 USDC on Base and take two working days once we agree on scope and have the samples. The results can guide any fine-tuning. Would that budget and payment method work for BAYTAS-INVA? Dima Mak / ub3dqy — Reply to this email directly, [view it on GitHub](#774?email_source=notifications&email_token=BWSGCVSI62SRCKQ4B7V7E3D5YTVFZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJVGUZDOOJVUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-19552795), or [unsubscribe](https://github.com/notifications/unsubscribe-auth/BWSGCVXAK3IDS4LBOP2XPC35YTVFZAVCNFSNUABIKJSXA33TNF2G64TZHMZTENRWGEYDONJXHNCGS43DOVZXG2LPNY5TCMJSG4YTKMJXUF3AE). Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for [iOS](https://github.com/notifications/mobile/ios/BWSGCVRHDNWQQQ5PKTJL7O35YTVFZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJVGUZDOOJVUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVJTG633UMVZF62LPOM) and [Android](https://github.com/notifications/mobile/android/BWSGCVX2CVE6JNNRUUI7NS35YTVFZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJVGUZDOOJVUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVZTG633UMVZF6YLOMRZG62LE). Download it today! You are receiving this because you authored the thread.

[ub3dqy]

Hi Jomart, yes. PP-OCRv5 Cyrillic mobile is already trained and ready to use offline. The 120 USDC pilot covers installing and testing the existing model on Windows with up to 10 anonymized scans and reference text. I would send setup steps, error rates by language, and processing times.

I can receive native USDC on Base. I suggest payment after I deliver the report and setup guide. In a wallet or exchange that supports Base, select USDC as the asset and Base as the network, then send 120 USDC to 0x6dD17C1D6CAd9D230F92bC193D0cF1D398e0CBD3. Share the transaction ID here and I can confirm receipt in writing. If your team needs a formal invoice, please let me know before we start.

If that works, let me know your Windows version and how many anonymized sample pages with reference text you have.

Dima Mak / ub3dqy

[SWHL]

Nice to see you working together in the RapidOCR community! Friendly heads-up: beware of scams. Looking forward to more collaboration!

[SWHL]

@baytasinva-dotcom

Q1: What is the recommended RapidOCR configuration for recognizing Kazakh, Russian, and English printed documents, including Kazakh-specific Cyrillic letters?

There are two existing offline models that support Cyrillic recognition. Please refer to the documentation for details: https://rapidai.github.io/RapidOCRDocs/latest/model_list/. You may test both of them.

You can also test the https://github.com/metrika-labs/PP-OCRv6_rec_cyrillic, and find the best effection model.

Q2: Can accuracy be further improved for clean A4 scans by fine-tuning or custom language data? We can provide representative sample scans.

You can refer to the PaddleOCR documentation, follow its steps to prepare datasets and configure training. It is not particularly difficult. After training is completed, you can deploy the model with RapidOCR.

Q3: Does the RapidOCR team or a developer you recommend offer paid customization? If so, what would be the approximate cost and development/testing timeline?

I’d suggest you check the PaddleOCR training documentation yourself. Training itself is not hard; the main challenge is whether you have sufficient training corpus and data.