WACP Advanced Hybrid Plagiarism & AI-Detection System
Ai Solutions
West African College of Physicians (WACP)
System Login
Staff and reviewers access the platform at wacpplagcheck.com through a secure login.
Overview
A multi-stage academic integrity platform for the West African College of Physicians, live at wacpplagcheck.com — combining n-gram fingerprinting, semantic embeddings, and a 4-class AI-content classifier into a single auditable pipeline for analysing dissertations, proposals, and reports.
The Problem
WACP's academic integrity reviewers were stuck with the same three failures common to conventional plagiarism checkers: re-uploaded files got flagged against themselves for a false 100% similarity score, resubmitted revisions (v1, v2, v3 of the same document) inflated similarity scores against their own earlier drafts, and — critically — none of the existing tools could detect ChatGPT, Gemini, Claude, or QuillBot-paraphrased text. Reviewers had no reliable way to trust a similarity score or catch AI-assisted submissions.
Our Solution
TeamO built a hybrid multi-stage detection pipeline: n-gram fingerprinting (Rabin-Karp rolling hashes) for exact sentence-level matches, sentence-transformer embeddings for paraphrase and near-duplicate detection, and sliding-window segmentation (150–300 word segments) for section-level precision. A `project_group_id` lineage system ensures resubmissions are never compared against their own earlier versions, and submissions are compared only against verified reference documents. A separate 4-class ML classifier distinguishes human-written, raw AI-generated, human-edited AI, and AI-paraphrased text, reporting results as probabilities (e.g. "Likely AI-paraphrased, Confidence: 71%") rather than absolute verdicts. The system runs a CodeIgniter 4 front end against a FastAPI/Celery/Redis analysis backend, with PDF.js-based reports that highlight matches directly on the original document — a full analytical report for admins, a highlighted summary (without internal model details) for students.
Technologies Used
Development Journey
Detection Dashboard
The dashboard shows project status, AI-detection distribution, and similarity severity across submissions.
Results
Live in production for WACP at wacpplagcheck.com, giving academic integrity reviewers auditable, explainable similarity and AI-authorship reports in place of the false-positive-prone output of conventional checkers. As of the latest dashboard snapshot, the platform has analysed 2,646 submissions with a 100% completion rate (zero pending or failed jobs) — 40 flagged for high similarity, and severity graded across 30,903 low, 38,039 medium, 21 high, and 19 critical segments. The AI-detection classifier has run against all 2,646 documents: 1,998 possibly human, 604 uncertain, 29 likely human, and 15 possibly AI-generated. Bulk processing handles 100+ submissions in 5–10 minutes via parallelised Celery workers and Redis-cached fingerprints/embeddings, versus roughly 50 minutes processed sequentially. The platform supports role-based access for admins and reviewing staff, English/French language switching, and a credit/subscription system for institutional usage.
Related Projects
Be the first to comment.