Transformer Fine-Tuning for Parental Control

Quick facts

/ CONTEXT

Traditional parental controls block entire websites. On mixed-content platforms, this also removes useful and educational material. I designed a smarter alternative: analyze content in real time and censor only inappropriate passages, applying generative AI and natural-language processing to a sensitive, real-world problem.

/ CHALLENGE

Precision was critical. An overly strict system becomes unusable, while a permissive one protects no one. It had to detect hate speech, offensive language, and obscene content accurately in Portuguese, then decide whether to blur a passage or block a page without disrupting browsing.

/ PROCESS

I fine-tuned BERTimbau, a Portuguese language model, for binary appropriate/inappropriate classification using HateBR, ToLD-BR, and OLID-BR. A FastAPI service returns a score from 0 to 1, while a Manifest V3 Chrome extension reads visible DOM text and sends it to the API.

/ SOLUTION

• Real-time analysis of visible page content.
• Selective blur for sensitive passages.
• Full-page blocking when harmful content exceeds the threshold.
• Configurable API URL, threshold, and guardian PIN.

/ HIGHLIGHTS

BERTimbau fine-tuned for Portuguese

Real time selective blurring

3 public datasets for toxicity

/ RESULTS

The project demonstrates that children can be protected online without sacrificing access to useful content. It produced a functional extension-plus-API architecture and strengthened my experience with deep learning, Transformer fine-tuning, APIs, and browser extensions.

/ TECHNICAL

The model is neuralmind/bert-base-portuguese-cased, fine-tuned with PyTorch and Hugging Face Transformers. The FastAPI backend supports individual and batch classification, and the JavaScript extension uses Service Workers and Manifest V3. BERTimbau was chosen for its stronger understanding of Portuguese nuance than generic multilingual alternatives.