Back to case studies

Case Study 06 · Full-Stack · Completed

Django Blog

A Django blog with rich-text and PDF posts, tags and search. Post excerpts come from a summarization model that runs locally with PyTorch and Hugging Face.

RoleSole author
Django Blog home page, styled as Operations Archive: crest logo, search bar, Add Documentation button and Archive, Purpose and Philosophy cards.
The home page puts search and adding a new post one click away.
Document library titled Knowledge Forge: a paged grid of six sample technical posts, each with an icon, title and generated excerpt.
Every post in the library carries an excerpt written by a local summarization model.
Post page for Deployment Checklist for Small Web Apps: DevOps and Docker tags, edit and delete buttons, a generated summary and PDF open and download.
PDF posts get a summary from the document's own text, plus open and download links.
Create New Documentation Post form: title, slug, TinyMCE editor with a sample article, PDF and featured-image uploads, tag checkboxes and linked posts.
Posts are written in a rich-text editor or uploaded as a PDF, then tagged.
Search results for django: six matching posts, the first showing its title, date, author, excerpt, Django, Python and Tutorial tags and Read More link.
Full-text search and tags make older notes easy to find again.

Problem

I wanted one place to keep and re-find technical notes, including PDFs. Writing an excerpt for every post by hand is the step that gets skipped, and without excerpts the post list is hard to scan.

Solution

A Django app with TinyMCE authoring, PDF upload, tags, full-text search, profiles and per-user comments. Excerpts are written automatically by a local summarization model, DistilBART through Hugging Face transformers on PyTorch, from either the post body or the uploaded PDF's text.

Highlights

  • The summarizer is a singleton, so the model loads once per process instead of once per request.
  • A distilled DistilBART checkpoint was chosen on purpose, trading a little quality for speed and memory.
  • Device selection is automatic: GPU when one is available, CPU otherwise, on the same code path.
  • PDF text is extracted with PyPDF2, so uploaded documents get excerpts too.
  • Comments pass through a profanity filter, and tags drive front-end filtering alongside search.

Outcome

Every post gets an excerpt with no extra work, and the model runs on the same machine as the app, with or without a GPU. Search and tag filtering make older notes easy to find again.

How the Work Ran

A solo project, built in small feature steps, with a later pass on speed and homepage caching.

Tech

  • Hugging Face Transformers
  • PyTorch
  • Django
  • Python
  • Tailwind
Case Study
Open full page