my projects

• cloudy (research project)

  viewgithub
  • early cloudburst predetermination system
  • signal decomposition: transforming the original signal using EEMD (Empirical Mode Decomposition) into multiple IMFs (Intrinsic Mode Functions)
  • feature extraction: by identifying different frequency components from each IMFs using FCR (Fine to coarse reconstruction)
  • integrated hardware for real-time data: ESP32 with DHT11 & beta rain sensor
  • data pre-processing: datetime transformations, cyclic/seasonal features, sliding window
  • prediction pipeline: LSTM, GRU, CNN-1D, & TFT (Temporal Fusion Transformer)
  • model evaluation: MSE, MAE, RMSE, FALSE NEGATIVE RATE, and F1-score
  • model re-training: DVC
  • NOT YET DEPLOYED......

  tech used: Python, Numpy, Pandas, Matplotlib, Seaborn, Scikit-learn, PyTorch, TensorFlow, Keras, HuggingFace Transformers, LangChain, Firebase

• adaptive rag for enterprise support systems (research project)

  viewgithub
  • low-latency, high-accuracy RAG system for enterprise query resolution
  • semantic search: SBERT embeddings + FAISS dense retrieval & vector store
  • contextual response generation: Llama3.2 2B with document re-ranking, & retrieval + hallucination grading
  • processed 1K+ agricultural/sustainable energy related docs & 100+ queries using spaCy & Tesseract OCR
  • outperformed baselines (like GPT-3.5, BART-only, TF-IDF, and BM25) on F1, ROUGE-L, BLEU, and Faithfulness
  • reduced response latency by 28% (~620 ms/query)
  • pipeline orchestration & re-training: LangChain, FastAPI, DVC
  • NOT YET DEPLOYED......

  tech used: Python, Pandas, spaCy, Tesseract OCR, PyTorch, HuggingFace Transformers, Sentence Transformers, ChromaDB, FAISS, LangChain, FastAPI, Docker, DVC