The archive · Developer & Business Tools · Product decision · 2024–2026
LLM-Aided OCR bets LLM cleanup can fix Tesseract; Show HN 479 pts, 3k GitHub stars
An MIT open pipeline runs Tesseract on scanned PDFs, then LLM-corrects and formats the text; its author sells document AI apps. 479 HN points, 3.0k stars.
LLM-Aided OCR
What the business is
LLM-Aided OCR is an MIT-licensed open-source pipeline that converts scanned PDFs to images, runs Tesseract, then uses an LLM (local via llama_cpp, or OpenAI/Anthropic APIs) to correct OCR errors, chunk text, format markdown and score output quality; its author also sells commercial web apps — including YoutubeTranscriptOptimizer.com — that apply the same document-AI approach.
How it started
A developer who goes by Dicklesworthstone released LLM-Aided OCR as an open-source project in August 2024, betting that large language models are the missing post-processing layer for OCR. The pipeline converts PDF pages to images, applies Tesseract with grayscale, thresholding and dilation preprocessing, splits the noisy text at sentence boundaries with overlap, and sends each chunk to an LLM for error correction and optional markdown formatting, with duplicate removal, header suppression and a quality assessment step.
What happened
The Show HN on 2024-08-09 drew 479 points and 172 comments. The project supported both local models (llama_cpp with custom grammars) and cloud APIs from OpenAI or Anthropic, with asynchronous chunk processing, adaptive token management, GPU acceleration and detailed logging. The README's closing section directed readers to the author's commercial web apps, including YoutubeTranscriptOptimizer.com, and a companion app that turns PDFs, Office files, images and audio into polished markdown documents with quizzes, slides, summaries and flashcards.
How it ended up
Still live: the repo recorded 3.0k stars, 214 forks and 64 commits at the 2026-09-04 crawl; no funding or company disclosures appear, and the author continues to use the open project as the front door to commercial document web apps.
Background
LLM-Aided OCR is an open-source project, released on GitHub in August 2024 by developer Dicklesworthstone, that treats large language models as the missing post-processing layer for OCR. The pipeline converts each PDF page to an image, runs Tesseract with grayscale, thresholding and dilation preprocessing, splits the raw text into sentence-boundary chunks with overlap, and sends each chunk to an LLM for error correction and optional markdown formatting.
The project supports local models through llama_cpp — including custom grammars for structured output — and cloud APIs from OpenAI or Anthropic, with asynchronous processing for API models, adaptive token budgeting, GPU acceleration and detailed logging. A quality-assessment step has the LLM score the corrected output against the raw OCR so users can see whether the fix actually worked. The launch Show HN on 2024-08-09 drew 479 points and 172 comments.
The README ends by pointing readers to the author's commercial web apps — including YoutubeTranscriptOptimizer.com and a companion app that converts PDFs, Office files, images and audio into polished markdown with quizzes, slides and flashcards — so the open repository functions as a working demo for the paid product line. By the 2026-09-04 crawl the repo carried 3.0k stars and 214 forks, with the project still maintained and MIT-licensed.
What has to be true
- Named-platform traction with a date: the 2024-08-09 Show HN (item 41203306) drew 479 points and 172 comments.
- A transferable bet: LLMs fix OCR errors in post — chunk, correct, format and score — turning decades-old Tesseract output into usable documents without retraining OCR models.
- Verifiable trajectory: the repo recorded 3.0k stars, 214 forks and 64 commits at the 2026-09-04 crawl.
- A visible funnel: the README's closing pitch routes open-source users toward the author's paid document web apps, making the free tool the demo for the business.
What can be applied
A free tool that visibly fixes the hard part — OCR garbage — can market a paid product to the exact users who need it, if the demo shows before/after quality rather than plumbing.
Aftermath
As of 2026-09-05 LLM-Aided OCR remains live as a maintained MIT-licensed open-source project: the 2026-09-04 crawl recorded 3.0k stars, 214 forks and 64 commits, with support for Tesseract plus local or API LLMs (OpenAI, Anthropic), markdown formatting, quality scoring and async processing. The README directs readers to the author's commercial web apps, including YoutubeTranscriptOptimizer.com and a companion document-to-markdown service, but no revenue, funding or company disclosures appear in the record — so the paid half of the funnel is unverified from this material.
Sources
- Show HN: LLM-aided OCR — Correcting Tesseract OCR errors with LLMs (479 points, 172 comments)
- Dicklesworthstone/llm_aided_ocr — repository (3.0k stars, 214 forks)
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card