AI Content Detector — Django + Machine Learning Final Year Project with Source Code

What This Project Actually Does

You paste some text into a box, or upload a .txt, .docx, or .pdf file, and the app tells you how likely it is that an AI wrote it. Not a vague "maybe" — you get a percentage, a verdict, and then the app goes sentence by sentence and highlights which specific lines look machine-written.

Behind that is a trained scikit-learn model. TF-IDF features plus stylometric signals (sentence length variance, punctuation patterns, vocabulary richness) fed into a calibrated Linear SVM. The model is already trained and bundled as three files, so you don't have to sit through hours of training on your laptop before the demo.

And the whole thing is one Django project. No separate Flask service, no Streamlit sidecar, no "run these two servers at once" nonsense that breaks in the lab five minutes before your viva.

Key Features

• Paste text or upload a document, get an AI-probability verdict in a couple of seconds
• Sentence-level highlighting — the suspicious lines get colour-coded so your examiner can see exactly what the model flagged
• Batch mode: drop in ten essays at once, download all results as a CSV
• Downloadable PDF report per scan, themed to match the app, ready to paste into your documentation
• Compare page that puts two texts side by side and shows stylometry differences
• Dashboard with Chart.js — scan volume over time, verdict mix, confidence trends
• REST API under /api/v1/ with X-API-Key authentication and per-key daily rate limits (hit the limit, you get a clean 429)
• A dark "Forensic Scan" UI built with Tailwind via CDN, so no npm build step to explain
• python manage.py seed_demo_data drops in 40 sample scans and a working demo API key — your dashboard looks alive from minute one

Real-World Applications

Colleges checking whether an assignment was written by ChatGPT. Content agencies verifying what freelance writers actually delivered. Editorial teams screening guest posts before publishing. Recruiters looking at cover letters that all sound suspiciously identical.

If you've already looked at a plagiarism checker project, this is the natural companion — plagiarism detection catches copied text, this catches generated text. Different problem, same academic-integrity shelf.

Who Should Buy This

If you're a BCA or MCA student in your final semester, your synopsis is already submitted with "machine learning" written on it, and you need something that actually runs — this is for you. It also works well for BTech CSE majors because there's enough depth here to survive a proper viva: feature engineering, model calibration, threshold tuning, API design, rate limiting, deployment.

Honestly, one thing is genuinely tricky. The model artifacts (tfidf_vectorizer.pkl, classifier.pkl, metadata.json) have to sit in exactly the right folder — detector/ml/artifacts/ — or the app throws an error on the first scan and you panic. That's it. That's the hard part. Run python manage.py check_model and it tells you straight away if the files loaded.

What's easier than it looks: the ML side. You're not training anything. The predictor loads the model once, caches it in memory, and everything after that is a normal Django view. If you can build a Django form, you can run this.

Why CodeAj

You get the complete source code — every view, every migration, every template, the test suite, the Gunicorn config, the systemd unit, the nginx file. A college-format project report comes with it, and if setup goes sideways we'll get on a call and fix it on your machine instead of sending you a link to Stack Overflow. Plenty of students end up browsing our Django projects with source code collection and our AI and ML final year projects hub before landing here, so if this one isn't quite the fit, there's more.

One more thing worth knowing before your viva. The bundled model reports around 99.98% accuracy on its held-out test split, and that number is real — but it's in-domain. Feed it a genre it never saw, or heavily human-edited AI text, and accuracy drops. Say that out loud in your presentation. Examiners love a student who knows the limits of their own model.