I'm a self-taught Data Scientist and AI Engineer based in Navi Mumbai, focused on building machine learning and AI systems that are reliable, explainable, and useful in real-world decision making.

My work spans predictive modeling, uncertainty quantification, retrieval-augmented generation (RAG), and memory-augmented AI systems.

• Probabilistic modeling — calibrated probabilities over hard classifications
• Uncertainty quantification — prediction intervals, confidence estimation
• Explainability — SHAP-based model transparency for regulated domains
• Time-series forecasting — demand forecasting with feature engineering
• Simulation — Monte Carlo methods for season-level uncertainty
• Applied AI systems — retrieval-augmented generation (RAG), semantic search, vector databases, and long-term memory architectures

• Retrieval-augmented generation (RAG) systems
• Long-term memory architectures
• Vector databases and semantic search
• AI evaluation and model routing
• Applied machine learning systems

End-to-end credit risk pipeline predicting loan default probability on the Home Credit dataset.

• Platt Scaling calibration reducing ECE from 0.346 → 0.001 — 99.7% improvement
• Risk bucketing (Low / Medium / High / Very High) aligned with lending policy
• SHAP explainability for individual applicant decisions and regulatory transparency
• Python Scikit-learn XGBoost SHAP

Conversational AI assistant featuring persistent memory, multi-session chat management, and real-time web retrieval.

• Four-layer memory architecture combining session memory, semantic vector retrieval, structured fact memory, and conversation summarization
• Automatic user profiling with semantic deduplication and memory conflict resolution to maintain accurate long-term user profiles
• Running conversation summaries to reduce prompt growth and preserve long-term context
• Intelligent model routing using lightweight and large LLMs to balance latency, cost, and response quality
• Real-time web search integration using Tavily, LangChain agents, and tool-augmented reasoning
• Streamlit deployment with caching, session persistence, and automated chat organization
• Centralized debug logging and modular memory architecture
• Python LangChain PostgreSQL pgvector Groq Streamlit

Hourly electricity demand forecasting on real EIA grid data (Texas, 2018–2023).

• Time-series feature engineering: lag features, rolling stats, cyclical encoding
• XGBoost achieving 2.40% MAPE — 48% improvement over seasonal naive baseline
• Weather integration via Open-Meteo API
• Python XGBoost Scikit-learn Pandas

Probabilistic match outcome modeling with explicit focus on draw modeling.

• Calibrated Home / Draw / Away probabilities using Platt Scaling
• Expected Points (xPts) league table from match-level probabilities
• 10,000 Monte Carlo season simulations for title, top-4, and relegation probabilities
• Python XGBoost Monte Carlo Simulation