Karthik

Mulugu

• Reduced document analysis time by 75% by building and deploying a production RAG system using LangGraph, Gemini, and FAISS/Pinecone with low-latency inference of 1.5–4.0s
• Improved retrieval precision by 30% by architecting a hybrid search pipeline combining semantic search, BM25, TF-IDF, and cross-encoder reranking — the core contribution of the IEEE Access paper

→ full benchmark results in the Research section

• Reduced stock price forecasting error by 49% (MAPE: 17.45% → 8.86%) by training and benchmarking Seq2Seq LSTM, LSTM, and Linear Regression models enhanced with RSI, MACD, and Bollinger Bands
• Delivered actionable insights across 6,300+ Amazon historical records (2000–2025) with an interactive dashboard integrating sentiment analysis and backtesting for data-driven buy/sell evaluation
• Deployed a production-grade pipeline on AWS EC2 — optimizing CPU-based PyTorch inference, enforcing secure network policies, and automating execution with systemd for reliable model serving

• Implemented VGG16 and ResNet50 from scratch in TensorFlow/Keras; then applied transfer learning with GoogLeNet and DenseNet to measure the accuracy-vs-compute tradeoff across all four
• ResNet50 with transfer learning achieved the best accuracy, outperforming the from-scratch VGG baseline by a significant margin on Stanford 40 Actions dataset
• Reached 80% real-time inference accuracy, with Streamlit app allowing image upload and live prediction — demonstrating viability for surveillance and fitness-tracking applications

• Fine-tuned facebook/bart-large-cnn on the CNN/DailyMail benchmark (300K+ article-summary pairs); achieved ROUGE-L of 0.41 — competitive with published abstractive baselines
• Abstractive output generates novel phrasings rather than copying source sentences — verified via n-gram novelty analysis against the reference summaries
• Streamlit interface accepts any URL or pasted text and returns a summary in under 3 seconds, validated across news, research abstracts, and legal documents

• Engineered a scikit-learn ColumnTransformer pipeline (imputation, scaling, encoding) with strict train/val/test isolation — eliminated 100% of null values and prevented data leakage across all 5 model families.
• Benchmarked Ridge, Random Forest, XGBoost, LightGBM, and PyTorch MLP with Optuna Bayesian tuning (30 trials/model); XGBoost achieved R2 = 0.969, RMSE = $2,035, MAE = $1,622 — 58% RMSE reduction over baseline.
• SHAP TreeExplainer identified Length of Stay, APR Severity Code, and Medical/Surgical pathway as the top three cost drivers; quantile regression (Q10/Q90) produced calibrated 80% prediction intervals.
• Deployed FastAPI REST API + 6-tab Streamlit dashboard (Predict, What-If, Batch, EDA, History) via Docker Compose; experiments tracked in MLflow.

• Fine-tuned distilbert-base-uncased on 204K HuffPost headlines with AdamW + linear warmup; converged at epoch 2 reaching 80.4% accuracy and 0.919 ROC-AUC — outperforming the BiLSTM baseline by 4.4 points
• Traced the BiLSTM's 76% ceiling to noisy category-based pseudo-labels: Business, Science, and Tech span the full sentiment spectrum, creating a data quality ceiling no architecture alone could overcome
• Built a live RSS pipeline fetching real-time headlines from 6 outlets (BBC, CNN, NPR), running batch sentiment prediction, and rendering hour-by-hour trend charts per source
• Deployed end-to-end: FastAPI inference backend, 4-tab Streamlit dashboard on Community Cloud, Docker for local deployment, MLflow experiment tracking, and GitHub Actions CI