A scanned Japanese toy manual is more than a block of text: diagrams, warnings, callouts, and button labels need to stay connected to the instructions they explain. Shy Ruparel’s Toku Solutions project addresses that problem by keeping each original page visible and placing editable English text over it. The engineering challenge was making the OCR, translation, and site-generation pipeline recover from failures without discarding completed work.
Why translating the words alone is not enough
Ruparel’s starting point was a collectible transformation toy and a manual he could not read. A plain-text translation could convey sentences, but it would not necessarily show which label belonged to a diagram, where a warning appeared, or which control an instruction described. For a multi-page illustrated manual, preserving those relationships is part of making the translation usable.
![Japanese Origami Kit for Kids: 92 Colorful Folding Papers and 12 Original Origami Projects for Hours of Creative Fun! [Origami Book with 12 projects]](https://m.media-amazon.com/images/I/61KOUlvb0TL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
Toku Solutions turns the PDF into a static site. Each page retains its original image, with English translated blocks positioned over it. Readers can edit those blocks and see the Japanese original, so OCR or translation errors can be corrected in context. Edits can then be sent back through GitHub as pull requests.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →How the manual becomes an editable site
- OCR: The system extracts text from the PDF, with work split into page-level Activities so pages can be processed in parallel.
- Translation: Extracted text is batched and translated. The translation Activity uses a Retry Policy configured for five attempts with exponential backoff.
- Cleanup: Deterministic encoding and rule-based cleanup run before Gemini is used for context-aware corrections.
- Site generation: The pipeline produces HTML and JSON for a static site served from a CDN. Longer rendering work uses heartbeating.
The result is not a replacement for the source document: the page image remains the visual reference, while the translated text is an editable layer. That combination makes it possible to inspect and correct questionable text without losing the manual’s layout.
#1 Best Overall
Why durable execution matters in this pipeline
OCR, translation, cleanup, and rendering are separate operations that can fail independently. If a later step fails, restarting the entire run would waste work already completed on earlier pages. Ruparel used Temporal Workflows to coordinate a parent Workflow and four Child Workflow phases: OCR, translation, site generation, and AI cleanup.
Temporal’s Event History records Activity executions, retries, and results. The command-line interface polls Workflow state instead of maintaining a separate database, Redis instance, or message queue as another source of state. In this design, workflow history and configured retry behavior help the pipeline resume around failures; they do not make every external service or task infallible.
Parallel work and retry behavior
Page-level OCR allows different pages to run in parallel. The translation stage batches extracted text and is configured to retry up to five times with exponential backoff. These choices address distinct issues: parallelism can reduce elapsed processing time, while retries can recover from transient failures. They do not guarantee that every page will succeed or that a retry will fix a bad OCR result.
Recommended Free Tools
Keeping workflow logic deterministic
The account places nondeterministic LLM work inside an Activity rather than deterministic Workflow code. That separation matters because an LLM response can vary, while workflow orchestration needs a reliable execution history. Pydantic validation checks that the cleanup response has the expected structure before the pipeline uses it.
Rank #3
- ❤ This diy miniature dollhouse kit includes all the furniture and accessories in the pictures.Japanese style dollhouse is suitable for the people over 15 years old.
- ❤ This mini house building kit are of high quality and the furniture and other wooden pieces are accurately cut to ensure they fit together correctly. LED lights,plants,decorations,furniture are quite realistic and cleverly designed.
- ❤ This DIY dollhouse kit mainly provide material packages and english manual, which require consumers to assemble themselves. Because the accessories are small,some friends may think that it will be difficult to assemble,You only need to follow the instructions step by step.when you finish it ,you'll be exciting.
- ❤ No one will refuse such a miniature dollhouse kit gift! If you are a DIY enthusiast,this dollhouse kits is perfect for you, more refined than ordinary crafts.
- ❤ The wooden dolls house has been 100% carefully checked before delivery to ensure that all accessories are complete and not missing. If you have any issues during assembly, please feel free to contact us directly, we will reply you as soon as possible. Note:Due to air freight restrictions on glue and batteries, You need to prepare by yourself.
Why AI cleanup is an enhancement, not the authority
Cleanup has deterministic encoding and rule-based stages before Gemini attempts context-aware corrections. If the LLM step fails after those earlier stages, the workflow can continue with partial cleanup. This avoids making Gemini a single point of failure, while accepting that the remaining text may need human review.
Ruparel describes the intended balance directly: “The AI is not pretending to be the final authority here.” The editable overlay and visible Japanese source are important for that reason: automation can produce a useful draft, but readers can correct mistakes where they occur.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the reported runtime and cost mean
In his 2026 account, Ruparel reports that a typical 20-page manual took about 50 seconds end to end and cost about 43 cents in his implementation. His approximate phase breakdown was:
| Phase | Reported time |
|---|---|
| Parallel OCR | About 30 seconds |
| Translation | About 5 seconds |
| AI cleanup | About 10 seconds |
| Static-site generation | About 5 seconds |
These are the author’s figures for his example pipeline, not a benchmark or a promise of typical performance. The account does not establish current API prices, translation-accuracy metrics, or that another manual, configuration, or service setup would produce the same results.
Best Value
- Used Book in Good Condition
When this approach fits—and when a simpler one may do
A screenshot translation feature may be convenient for a single page or an occasional lookup. A pipeline that extracts, translates, and generates editable pages is more relevant when the goal is to process a multi-page manual, keep the layout available for reference, and share corrections. The trade-off is the engineering required to build and operate the pipeline, including OCR, translation, validation, and failure handling.
Likewise, deterministic cleanup is repeatable and easier to validate, while LLM cleanup can make context-aware adjustments but needs structured-output validation and a fallback path. Ruparel’s design combines them rather than treating either as sufficient alone. His account describes one implementation, not a controlled comparison with other orchestration products or a measured evaluation of translation quality.
Source and scope
This account is based on Shy Ruparel’s article for Temporal, republished on DEV Community on July 20, 2026: “Why I needed Durable Execution to read a toy manual”. The article describes the Bandai CSM collectible transformation-toy context but does not identify the exact model; the workflow itself does not require owning a toy.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
