http://localhost:1313/posts/ Recent content in Writing on Divyam Ahuja Hugo en-us Fri, 08 May 2026 00:00:00 +0000 http://localhost:1313/posts/hello-world/ Fri, 08 May 2026 00:00:00 +0000 http://localhost:1313/posts/hello-world/ <p>Welcome to my new blog! I&rsquo;ve rebuilt my personal website from scratch using <a href="https://gohugo.io/">Hugo</a>, a fast static site generator written in Go.</p> <h2 id="why-hugo">Why Hugo?</h2> <p>I wanted something minimal, fast, and that gets out of the way. Hugo checks all the boxes:</p> <ul> <li><strong>Blazing fast builds</strong> — the entire site builds in under 100ms</li> <li><strong>Markdown-first</strong> — I write posts in plain markdown files</li> <li><strong>Zero JavaScript</strong> — the site ships pure HTML and CSS</li> <li><strong>Built-in features</strong> — syntax highlighting, RSS feeds, sitemaps, all out of the box</li> </ul> <h2 id="the-design">The Design</h2> <p>The design is inspired by the <a href="https://probberechts.github.io/hexo-theme-cactus/">Cactus Dark</a> theme — a minimalist, terminal-inspired aesthetic with monospace typography and a dark color scheme. I built the theme from scratch for full control.</p> http://localhost:1313/posts/pdf-translation-pipeline/ Wed, 11 Feb 2026 00:00:00 +0000 http://localhost:1313/posts/pdf-translation-pipeline/ <p>Recently, I needed to translate a PDF document from Hindi to English. Sounds simple enough, right? Turns out, it&rsquo;s a surprisingly deep rabbit hole.</p> <h2 id="the-pipeline">The Pipeline</h2> <p>The approach I settled on follows this flow:</p> <div class="highlight"><pre tabindex="0" class="chroma"><code class="language-fallback" data-lang="fallback"><span class="line"><span class="cl">PDF → Images → OCR → Translation → Rendered Images → PDF </span></span></code></pre></div><p>Each step has its own set of challenges:</p> <ol> <li><strong>PDF to Images</strong>: Convert each page to a high-DPI image for better OCR accuracy</li> <li><strong>OCR</strong>: Extract text with position data using PaddleOCR</li> <li><strong>Translation</strong>: Run extracted text through NLLB (No Language Left Behind)</li> <li><strong>Rendering</strong>: Paint translated text back onto the original image</li> <li><strong>Assembly</strong>: Combine rendered images back into a PDF</li> </ol> <h2 id="lessons-learned">Lessons Learned</h2> <ul> <li><strong>DPI matters a lot</strong> — bumping from 150 to 300 DPI dramatically improved OCR accuracy for Hindi text</li> <li><strong>Font rendering is hard</strong> — getting translated text to fit in the same bounding boxes required careful font size calculation</li> <li><strong>Fallback strategies</strong> — TrOCR as a fallback when PaddleOCR fails on certain text regions</li> </ul> <p>The code is messy but it works. Sometimes that&rsquo;s enough.</p>