Moulya Reddygari Bhupal

Production-grade Retrieval-Augmented Generation system with modular, role-based architecture for grounded interview workflows.

• Built with LangChain, LangGraph, and ChromaDB; top-k semantic retrieval with context injection and a hallucination guard for grounded responses.
• Scalable document ingestion with duplicate detection, recursive chunking (100–300 words), metadata tagging, and embedding-based indexing.
• LangGraph interview assistant with multi-stage prompts for question generation, answer evaluation, and structured responses.
• Source-grounded outputs with structured JSON and citation support for traceability.

Tools used: Python · LangChain · LangGraph · ChromaDB · Groq · LLMs · Streamlit · vector databases

Time-series forecasting on 1M+ records with 34 features using predictive modeling for smart-grid energy use.

• Achieved R² 0.969, RMSE 94.33, MAE 66.03 with LightGBM after extensive feature engineering and model comparison.
• Engineered 20+ features: cyclical encoding, lags, rolling statistics, weather indicators (HDD/CDD).
• Compared Linear, Ridge, Random Forest, XGBoost, LightGBM, ANN, with ensemble improvements; time-series split validation to prevent leakage.
• Custom ETL for multi-source data, KNN imputation, and IQR-based outlier handling.

Tools used: Python · Pandas · scikit-learn · XGBoost · LightGBM · time-series forecasting · feature engineering

ML pipeline on the NYC TLC dataset to predict fares from trip-level and engineered distance and temporal features.

• Compared Linear Regression, Decision Tree, and Random Forest; feature engineering drove the largest gains.
• Preprocessing: missing-value imputation and outlier removal for higher data quality.
• Streamlit app for real-time fare prediction and interactive visualization.

Tools used: Python · Pandas · scikit-learn · Streamlit · feature engineering

End-to-end analytics platform ingesting data from APIs (e.g., World Bank, Google Trends) and visualizing real-time metrics across 150+ countries.

• Designed Python/SQL pipelines for automated ingestion, transformation, and storage in PostgreSQL with strong data-quality controls.
• Built interactive Tableau and Streamlit dashboards for regional analysis, metric selection, and trends—with query times under about 2 seconds; evaluated Apache Spark for large-scale multi-country ingestion.
• Automated refresh with Airflow and Google Cloud Scheduler, cutting manual reporting effort by roughly 60%.

Tools used: Python · SQL · Pandas · PostgreSQL · Streamlit · Airflow · Tableau · Google Cloud Scheduler

Reinforcement-style game agent that learns to play 2048 from experience, using reward feedback to improve policy over time.

• Trained a deep neural network to replace a human player from game state, improving through continued gameplay.
• Outperformed rule-based baselines by about 35%; reached scores of 65,000+ and the 2048 tile within ~25 minutes of learning.

Tools used: Python · TensorFlow · Keras · Pandas · NumPy

Relational database for an event-management catering workflow—food supply chain, events, staff, and payroll—with less redundancy and clearer operations.

• MySQL schema and processes that reduced data redundancy by about 35% and manual record-keeping effort by about 40%.
• Led a team of four with on-time delivery; schema designed with scalability as data volume grows.
• Visualization and analysis for stakeholders; earned Best Presentation Award and a professor recommendation letter.

Tools used: MySQL · Python

Hybrid recommender combining collaborative filtering with deep learning for personalized product suggestions on simulated e-commerce data.

• Matrix factorization plus neural components; ~25% higher recommendation accuracy on a 10,000-user synthetic dataset vs. baselines.
• Published in IEEE ICRTDA 2024 proceedings; presented at NCKITS 2023 (Tirupati, India).

Tools used: Python · Pandas · NumPy · TensorFlow · Keras · scikit-learn