Dean

@codevore1

building opensource Document Intelligence @ docuglean-ai. Documents → structured data http://git.new/docuglean-ocr RLHF. RL Envs.

git.new/docuglean-ocr

Joined June 2018

331Posts 297Followers 472Following

You might like

@iamaagam1

@uxmohamed_

@derekisnt

@gabrielwllnx

@somilkumar2470

@ChiehYun6

@240p60

Pinned

Dean

@codevore1

Sep 2

Excited to announce I’m building Docuglean - an open-source, privacy-focused document intelligence layer with three key foundations:. • Agentic OCR for parsing documents into high-quality, ready-to-use data. powered by SOTA VLM models for complex tables, formats, and layouts.…

codevore1's tweet image. Excited to announce I’m building Docuglean - an open-source, privacy-focused document intelligence layer with three key foundations:.

• Agentic OCR for parsing documents into high-quality, ready-to-use data. powered by SOTA VLM models for complex tables, formats, and layouts.…

Dean

@codevore1

Sep 23

Love the huge OCR boost. Qwen3VL-235B-A22B beats GPT5 and Claude-Opus 4.1

Qwen

@Alibaba_Qwen

Sep 23

🚀 We're thrilled to unveil Qwen3-VL — the most powerful vision-language model in the Qwen series yet! 🔥 The flagship model Qwen3-VL-235B-A22B is now open-sourced and available in both Instruct and Thinking versions: ✅ Instruct outperforms Gemini 2.5 Pro on key vision…

Alibaba_Qwen's tweet image. 🚀 We're thrilled to unveil Qwen3-VL — the most powerful vision-language model in the Qwen series yet!

🔥 The flagship model Qwen3-VL-235B-A22B is now open-sourced and available in both Instruct and Thinking versions:
✅ Instruct outperforms Gemini 2.5 Pro on key vision…

Dean

@codevore1

Sep 18

used bm25 for a doc search app and it was way better. faster, cheaper and gives results the way a user would search for docs. chose bm25s since it was smaller and known for its extremely fast retrieval with nicer results vs Elasticsearch, BM25-PT and Rank-BM25

codevore1's tweet image. used bm25 for a doc search app and it was way better. faster, cheaper and gives results the way a user would search for docs.

chose bm25s since it was smaller and known for its extremely fast retrieval with nicer results vs Elasticsearch, BM25-PT and Rank-BM25

Dean

@codevore1

Sep 16

VGT (GiT and ViT) + a specialized OCR engine like PaddleOCR and a VLM like Qwen for fallbacks is all you need for a high-quality structured data ready for LLMs Building this for our agentic ocr and def looking for testers. Send me your most complex PDFs :)

codevore1's tweet image. VGT (GiT and ViT) + a specialized OCR engine like PaddleOCR and a VLM like Qwen for fallbacks is all you need for a high-quality structured data ready for LLMs

Building this for our agentic ocr and def looking for testers. Send me your most complex PDFs :)

Dean

@codevore1

Sep 15

Oldie but goodie. Stumbled across this late 2024 paper while digging into recent OCR research. The timing works well with the advanced doc layout analysis solutions I've been working on

codevore1's tweet image. Oldie but goodie.

Stumbled across this late 2024 paper while digging into recent OCR research. The timing works well with the advanced doc layout analysis solutions I've been working on

Dean

@codevore1

Sep 11

uv + ruff is all you need

Dean

@codevore1

Sep 10

reading this research paper on using Vision Grid Transformer for Document Layout Analysis. Pretty solid approach for breaking down document pages into structured components like headers, titles, and body text. The value lies when you extract text from each section separately and…

codevore1's tweet image. reading this research paper on using Vision Grid Transformer for Document Layout Analysis. Pretty solid approach for breaking down document pages into structured components like headers, titles, and body text. The value lies when you extract text from each section separately and…

Dean reposted

Dean

@codevore1

Sep 2

Dean

@codevore1

Aug 27

I reviewed 100 models over the past 30 days. Here are 5 things I learnt. TL;DR: Spent a month testing every AI model for work, a few tools I'm building and RL. Build task-specific evals. Most are overhyped, a few are gems, model moats are ephemeral, and routers/gateways are the…