🔬 Technical Overview

The system is a PHP-driven application that orchestrates custom-trained transformer models for English ↔ Swahili translation. Language identification is handled first via a K-Nearest Neighbor classifier, and the detected text is then forwarded to the transformer for translation. The pipeline is optimized for low-resource conditions, ensuring dependable inference even on modest infrastructure - runs on either CPU or GPU enviroments

🏗️ Architecture & Training

The model extends Vaswani et al.’s transformer - Attention Is All You Need - with targeted adjustments for Swahili morphology:

• Multi-objective loss (cross-entropy, unlikelihood)
• Attention head rebalancing plus differentiated dropout schedules
• Subword tokenization tuned for noun classes and agglutinative patterns
• Weight tying between embeddings and output projection layers
• Beam search with n-gram blocking and semantic penalties

🧱 Platform Stack

• Backend: PHP + custom ML services with optimized memory pools
• Frontend: Vanilla HTML/CSS/JS with progressive enhancement
• Inference: Transformer checkpoints served via lightweight PHP bindings
• Observability: Confidence scoring, request tracing, GPU safety checks

🤝 Collaboration

We welcome dataset contributions, evaluation feedback, and experiments around Swahili dialect coverage. Reach out if you are exploring transfer learning, domain adaptation, or cultural-linguistic UX studies and general feedback