Data Scientist · Machine Learning & Generative AI
Data Scientist building end-to-end machine learning and generative AI systems — from research and modelling through to deployment and inference optimization. My work spans large language models (LLMs), predictive modelling, and time-series forecasting, with a focus on turning applied ML into reliable, production-ready tools.
- 🏥 Data Scientist (R&D) at Innovaccer (US healthcare), since 2024
- 🎓 B.E. Computer Science, Chitkara University (CGPA 9.3/10)
- 🔬 Interests: LLM fine-tuning & RAG, forecasting, MLOps, applied research
- Time-series forecasting — a versatile forecasting engine for multiple KPIs using SARIMAX, Exponential Smoothing and Prophet, with Bayesian optimization / HyperOpt tuning and SMAPE-based model selection (sustained SMAPE of 1–3 across KPIs).
- Agentic GenAI — an in-house Retrieval-Augmented Generation (RAG) Q&A assistant with multi-agent orchestration (LangChain, CrewAI) over mixed data (text, video, databases), backed by FAISS/ChromaDB vector search and hybrid semantic retrieval.
- Predictive modelling & data engineering — churn-risk and attribution modelling for value-based care (CareConnect): ETL pipelines (Python/SQL) into Snowflake and tuned models (+12% AUC, −40% training/inference time via multiprocessing).
- LLM-powered automation — GPT-4 tooling such as DAX→SQL conversion + validation on Snowflake, and a codebase dependency/lineage roadmap generator across 500+ SQL/Python files.
| Project | Highlights | Tech | Links |
|---|---|---|---|
| Hospital-KPI Forecasting Engine | Multi-KPI US-healthcare demand forecasting — deep EDA, SARIMAX + Fourier / Holt-Winters / Prophet, rolling-CV SMAPE selection (~1.4% on operational KPIs). | statsmodels · Prophet · Streamlit |
📊 Slides |
| Automated Valuation Model (AVM) | Explainable, uncertainty-aware property valuation — LightGBM + Optuna, SHAP explanations, conformal prediction intervals (R² 0.85). | LightGBM · SHAP · Streamlit |
📊 Slides |
| Customer Churn Prediction | Behavioural churn on UCI Online Retail II — leakage-free time-split, RFM features, 5-model comparison (Gradient Boosting, ROC-AUC 0.80). | scikit-learn · SHAP · Streamlit |
📊 Slides |
📊 Slides open in your browser (Office viewer) ·
Languages
ML / Deep Learning
Generative AI / NLP
Data / Cloud
Tools
Open to opportunities in Data Science, Machine Learning and Generative AI.
📫 Reach me on LinkedIn or at manikkaura2002@gmail.com