Overview
A Turkish-learning assistant for foreigners: I fine-tuned BERTurk and mT5 for grammar error detection and pedagogical correction and built the multi-platform Flutter app.
Technical Approach
- BERTurk fine-tuned for binary classification on a custom dataset built from TrCoLA and opus100.
- mT5-small trained on wrong→correct sentence pairs (seq2seq) to produce corrections and pedagogical explanations.
- BERTurk (classification) and mT5 (seq2seq correction) combined into a single pipeline behind a FastAPI REST API.
Key Features
- Multi-platform Flutter app for iOS, Android and desktop
- Scenario-based chatbot (tea garden, market, pharmacy…)
- Sentence corrector and progress-tracking modules
System Architecture
Client
FlutterChatbot · Corrector · Progress
API
FastAPIREST
Models
BERTurk (classification)mT5-small (seq2seq)
Results & Achievements
73%
Test accuracy (BERTurk)
0.73
Macro F1
Gallery
05 / Contact
Let’s build something intelligent.
Interested in autonomous systems, software or ambitious engineering projects? I’d love to hear from you.