Shipped a live Sinhala AI-text detector with explainable predictions

SinXdetect · Final Year Research Project · 2025 – 2026

Context

Sinhala-language academic and media content is increasingly AI-generated, but existing detectors are English-first and unusable for Sinhala script.

Problem

Educators need a verdict they can trust and defend. A raw probability score is not enough — reviewers need to see which words drove the classification.

My role

Product owner and lead engineer for a final-year research project: problem framing, dataset decisions, model evaluation, application build and deployment.

Constraints

Limited labelled Sinhala corpus
Single GPU budget for fine-tuning
Explanations had to be fast enough for interactive use

What I did

My contribution

Framed the product around explainability rather than a single confidence score
Fine-tuned a SinBERT transformer for human vs AI binary classification
Added LIME-based word-level explanations to every prediction
Built the FastAPI backend with batch processing and the React + Tailwind UI
Containerised the stack with Docker Compose and deployed to sinxdetect.movindu.com

Owned by the wider team

Supervisor provided research direction and evaluation review

Outcome

Live, publicly usable detector at sinxdetect.movindu.comHow this was measured: Verified through the publicly accessible deployment

What I'd do differently

I would validate the explanation UI with real educators before tuning the model further — trust was the actual product problem.
I would version datasets formally from the first experiment.