Karthik

Mulugu

• Reduced RMSE by 58% over the linear baseline by benchmarking 5 model families with Optuna hyperparameter search, achieving R²=0.969 and RMSE=$2,035 on 2.5M NY state inpatient records with XGBoost.
• Enabled clinicians to act on uncertainty rather than point estimates by combining SHAP TreeExplainer for feature attribution with quantile regression for calibrated 80% prediction intervals, resulting in interpretable and trustworthy cost forecasts.
• Reduced deployment to a single command by packaging a FastAPI inference service, 6-tab Streamlit dashboard, and MLflow experiment tracker into Docker Compose, resulting in a fully reproducible one-click production environment.

• Achieved 75% faster document analysis by orchestrating summarisation, comparison, refinement, and QA modes through LangGraph multi-intent routing backed by Gemini, eliminating manual query-type switching.
• Improved retrieval precision by 30% over pure dense retrieval by implementing hybrid BM25 + FAISS + TF-IDF with cross-encoder reranking, resulting in higher factual QA accuracy on complex multi-page PDFs.
• Identified a critical failure mode in fixed-weight hybrid fusion by benchmarking lexical vs. dense retrieval across diverse query types, resulting in a novel query-aware fusion gating approach submitted to IEEE Access.

→ see Research section for full paper details

• Outperformed BiLSTM by 4.4 accuracy points by fine-tuning DistilBERT on 204K HuffPost headlines with noise-robust multi-class labeling, achieving 80.4% accuracy and 0.919 ROC-AUC.
• Enabled real-time public-sentiment monitoring across 6 major outlets (BBC, Reuters, CNN, NYT, Guardian, AP) by building a live RSS ingestion pipeline with hour-by-hour trend visualisation, resulting in a continuously-updated news analytics dashboard.
• Reduced deployment to a single command by packaging FastAPI inference, Streamlit UI, and MLflow tracking into Docker Compose with GitHub Actions CI, resulting in automated continuous deployment to Streamlit Cloud on every push.

• Reduced Amazon stock MAPE by 49% (17.45%→8.86%) by training a Seq2Seq LSTM with encoder-decoder architecture on 6,300+ daily records fused with RSI, MACD, and Bollinger Band indicators, resulting in materially sharper price forecasts.
• Improved buy/sell signal quality by integrating a financial news sentiment pipeline with price signals and backtesting against the S&P 500 benchmark, resulting in data-driven trade evaluation beyond price-only models.
• Ensured continuous model freshness by automating data refresh and retraining on AWS EC2 with systemd, resulting in a live Streamlit dashboard that reflects the latest market data without manual intervention.

• Achieved 88% recall on the minority churn class by training XGBoost on 7,000+ telecom records with 12 SQL-engineered features, optimising for the metric that directly drives retention campaign ROI.
• Identified that 63% of churn originates from high-charge month-to-month customers by running SQL cohort analysis, resulting in a clear segment-first targeting strategy that reduces campaign spend waste.
• Enabled non-technical stakeholders to act on churn predictions without analyst support by deploying a Streamlit app with a live threshold slider and per-customer SHAP waterfall charts, resulting in self-serve retention decisions.

• Achieved best accuracy with 3× fewer training epochs by benchmarking ImageNet transfer learning (GoogLeNet, DenseNet) against training-from-scratch (VGG16, ResNet50) on 9,500+ Stanford 40 images, confirming DenseNet transfer as the optimal architecture.
• Delivered 80% real-time inference accuracy on held-out test data by deploying the best model through a live Streamlit app supporting webcam and image upload, resulting in an interactive action classification demo.
• Produced reproducible training insights by running a systematic ablation study over batch sizes, learning rates, and augmentation strategies across all four architectures, resulting in documented loss/accuracy curves for each configuration.

• Eliminated manual quality review by building a 4-step agentic pipeline (domain detection → strategy selection → ROUGE-L evaluation → auto-retry) with Groq Llama 3.3 70B, resulting in self-correcting summaries that meet quality thresholds without human intervention.
• Reduced perceived latency to zero by streaming tokens in real time via SSE while supporting text, PDF, and URL inputs with TXT/JSON/Markdown export, resulting in a production-grade tool built entirely on a single Vanilla JS frontend.
• Achieved zero-downtime continuous releases by configuring GitHub Actions to containerise via Docker and push to HuggingFace Spaces on every commit, resulting in a production environment that is always up to date.