Member-only story

I Tested Chandra OCR: Converting Handwritten & Scanned Documents into Structured, Layout-Aware HTML

A hands-on guide to building a GPU-powered OCR pipeline that preserves layout, tables, math, and handwriting in clean HTML.

7 min readMar 2, 2026

👨🏾‍💻 GitHub ⭐️ | 👔 LinkedIn | 📝 Medium | ☕️ Ko-fi | 🌐 InventaAI Blog

Introduction

I recently spent time integrating Chandra OCR into a scan, a handwritten, lab report with a complex table and formula, included in the documents pipeline running locally with an NVIDIA A10G GPU. What I found was something genuinely different: a model that outputs not just text, but structured HTML with bounding boxes, layout labels, tables with proper colspan and rowspan, inline math, and checkbox states. I have a build pipeline for a scanned lab report with merged table cells, a handwritten note, a financial form with checkboxes, and a research paper full of LaTeX equations. Traditional OCR tools, such as Tesseract, PaddleOCR, and even Google Vision, treat documents as bags of text. They read words, but they don’t understand pages.

This post walks through everything: what Chandra is, how it works, how to install and run it on your own GPU, and how to reconstruct clean Word documents from its output.

Data Science Collective

I Tested Chandra OCR: Converting Handwritten & Scanned Documents into Structured, Layout-Aware HTML

A hands-on guide to building a GPU-powered OCR pipeline that preserves layout, tables, math, and handwriting in clean HTML.

Introduction

Written by Md Monsur ali

Published in Data Science Collective

Responses (1)

More from Md Monsur ali and Data Science Collective

Huawei’s $2000 GPU with 96GB VS Nvidia — Is that the end of the NVIDIA Monopoly?

Huawei’s Atlas 300I Duo is real, it’s cheap, and it has more VRAM than most of us will ever own. That does not make it an Nvidia killer…

Is Running a Local LLM Still Worth It After Haiku 5.5?

A cheap cloud model now matches Qwen 3.8 27B. What that does to the 5 usual reasons for running a model on your own GPU.

What Is the Best Local LLM for Mac in 2026? (M1 to M6)

A 2021 M1 Pro writes answers faster than a new M5 Air. The best model for every Mac from 8GB to 512GB, plus voice, embeddings and image.

Generative AI vs AI Agents vs Agentic AI: Key Differences Explained with Examples

The core distinctions between Generative AI, AI Agents, and Agentic AI, with real-world use cases, examples, and when to use each in modern…

Recommended from Medium

I Replaced My $200/Month Claude Code with These Local Models.

One model for the agent, one for coding, one for autocomplete, and what fits on 24GB, 48GB and 64GB Macs.

Building a RAG Pipeline for 10M+ Documents With Near-Zero Hallucination

Retrieve, constrain, verify, abstain

I Run Qwen3.8–27B as My Local Personal Agent — And It Feels Amazing

I have been running local LLMs since the Llama 2 days. Most were fun toys. Qwen3.8–27B is not a toy. It is the first local model that feels…

OCR Showdown — Evaluating Three OCR Models

Deciding which OCR model is the best for day-to-day usage.

Run Jev Decision Models on Your Laptop With Ollama

Ollama 0.35 added a decision endpoint that turns small local models into typed answering machines. Here is the 10-minute setup.

Meet terrio: Geospatial Data I/O in Pure Rust, No GDAL Required

Rasters, vectors and point clouds from one CLI, Python, JavaScript and Rust | Runs in the browser | An MCP server for AI agents 🦀🌍