EKEi Ei Khaing
Contact
HomeProjectsExperienceEducationContact

Projects

Selected projects across full-stack delivery and applied AI/ML.

Smart Travel System

AI-powered trip planning platform using LLMs + real-world APIs for itineraries, sentiment insights, and language support.

Next.jsReactTypeScriptFastAPIPostgreSQLGemini LLMGoogle Places API
  • Designed a modular backend architecture that separates API, services, and AI orchestration layers to support scalability and future feature expansion.
  • Integrated Gemini (2.5 Flash) for itinerary generation, travel Q&A, and language buddy features.
  • Built REST APIs with FastAPI + PostgreSQL and connected end-to-end UI → API → AI output.
  • Implemented review-based sentiment insights from destination data to support better decisions.

Bike Sharing Demand Forecasting

End-to-end time-series forecasting with GRU + Optuna optimization + SHAP explainability.

PythonTensorFlow/KerasOptunaSHAPPandasNumPy
  • Built hourly forecasting pipeline with lag features and time encodings, evaluated with RMSE/MAE/R²/sMAPE.
  • Optimized hyperparameters using Optuna (Random Search + TPE) and compared performance across trials.
  • Explained global + local predictions using SHAP for interpretability and trust.

Fashion Product Image Classification (CNN vs ViT)

Computer vision comparison of CNN and Vision Transformer architectures in TensorFlow.

PythonTensorFlow/KerasCNNVision Transformer (ViT)
  • Implemented custom CNN and ViT models and compared accuracy, training efficiency, and generalization.
  • Created preprocessing + augmentation workflow and analyzed trade-offs between architectures.
  • Demonstrated where attention-based models can outperform CNNs on subtle category differences.

English → French Machine Translation

NLP translation using Seq2Seq and Transformer approaches with BLEU evaluation.

PythonSeq2Seq (GRU/LSTM)TransformersHuggingFaceBLEU
  • Built Seq2Seq translation baselines and experimented with Transformer-based approaches.
  • Evaluated translation quality using BLEU and performed error review.
  • Benchmarked against pre-trained HuggingFace translation models for comparison.