In this study, we present the development and evaluation of an online IELTS practice system integrated with automatic scoring modules for the Writing and Speaking skills. The system is designed for practice purposes, allowing learners to submit responses, receive criterion-based predicted band scores, and monitor score changes across multiple practice attempts, rather than replacing human examiners in official assessment contexts. For the Writing module, the prompts and responses are encoded using all-MiniLM-L6-v2 and then combined with 52 linguistic indicators at a fusion layer to predict four criteria: Task Response/Task Achievement (TR/TA), Coherence and Cohesion (CC), Lexical Resource (LR), and Grammatical Range and Accuracy (GRA). For the Speaking module, each sample is represented by 420 multi-source features extracted from transcripts, semantic representations, fluency-related signals, and selected acoustic features; four XGBoost regressors are used to predict Fluency and Coherence (FC), Lexical Resource (LR), Grammatical Range and Accuracy (GRA), and Pronunciation (PR). Experiments were conducted on 2,162 writing responses and 281 speaking samples. The results show that the Writing module achieved an overall Quadratic Weighted Kappa (QWK) of 0.6482 on the test set after averaging the outputs of three models, while the Speaking module obtained an average QWK of 0.6240 under 5-fold cross-validation. These preliminary findings indicate that integrating automatic scoring models into an online practice platform is feasible. However, the Speaking dataset remains small, the labelling process has not yet been fully standardized, and the system has not generated qualitative feedback directly linked to the IELTS rubric. Therefore, future work should focus on expanding the dataset, controlling data leakage, standardizing labels, and incorporating qualitative feedback mechanisms before large-scale deployment.