Projects
Below is a selection of projects that reflect my work in predictive modeling, LLMs, simulation, and applied machine learning for energy, climate, and public health. Each project emphasizes decision relevance alongside technical rigor.
⚡ CAISO Day-Ahead Electricity Price Forecasting Pipeline (Side Project)
Focus: Energy forecasting · MLOps · Machine learning
Developed an end-to-end, configurable forecasting pipeline for day-ahead electricity prices in the California ISO (CAISO) market. The project was designed to mirror real market operations by generating hourly forecasts prior to the 9 AM day-ahead bidding deadline while enforcing leakage-safe feature engineering and reproducible MLOps practices.
Key contributions:
- Designed a modular forecasting pipeline integrating historical electricity prices with demand, renewable generation, and fuel price data
- Developed leakage-aware feature engineering workflows that correctly handle forecast issue time, time zones, and daylight saving time transitions
- Implemented configurable training and inference pipelines with automated hyperparameter tuning, tree-based feature selection, and model registry integration through MLflow
- Benchmarked LightGBM and XGBoost against multiple persistence-based baselines and analyzed model behavior using quantitative metrics and qualitative forecast comparisons
- Designed the pipeline to support multiple forecasting strategies, including price-only, exogenous-only, and hybrid models through configuration rather than code changes
Methods & tools:
Python, LightGBM, XGBoost, MLflow, scikit-learn, DuckDB, Pandas, YAML, Time-Series Forecasting
Outcome & impact:
- Built a reproducible forecasting pipeline following MLOps best practices from data ingestion through model deployment
- Demonstrated that combining historical prices with demand, renewable generation, and fuel price features consistently outperformed price-only and exogenous-only approaches
- Established a flexible foundation for future extensions including probabilistic forecasting, deep learning (LSTM/TFT), and optimization of battery energy storage systems (BESS) and day-ahead/real-time trading strategies
⚡ Northeast Freight Corridor Charging Plan
Focus: Predictive modeling · Simulation · Energy systems
Developed forecasting and simulation-based models to support proactive grid infrastructure planning under increasing electrification from EVs.
Key contributions:
- Designed, led, and deployed predictive simulation models, with uncertainty quantification enabled, using Geotab’s vehicle telematics data across 140 highway plazas in 9 Northeast states, enabling National Grid and state utilities to plan multi-megawatt medium- and heavy-duty charging infrastructure in support of a national freight electrification strategy
- Developed a modular system to prepare, process, and visualize data for scalable, reusable analytics
- Co-led the development of a client-facing electrification roadmap, synthesizing technical analysis into actionable policy and investment guidance for utilities, regulators, and state agencies
- Co-led workshops to communicate the insights of the project to various stakeholders, ranging from utilities to state representatives in the Northeast
Methods & tools:
Python, Monte Carlo Simulation, Vehicle Telematics Data, Forecasting, Datawrapper
Outcome & Impact:
- Enabled planners to assess grid readiness
- Located areas that need prioritized investment
- Identified potential constraints
- Compared investment strategies under uncertainty
🔗 Project Link 1 🔗 Project Link 2 🔗 Full Report Link 🔗 Summary Link
🧠 EVRAG — Large Language Models for Interpretation (Side Project)
Focus: LLMs · RAG · Decision support
Developed a Retrieval-Augmented Generation (RAG) system to assist with interpreting and communicating insights from technical EV charging reports.
This project treats language models as a supporting layer — augmenting quantitative modeling rather than replacing it.
Key contributions:
- Built a modular RAG pipeline grounded in domain-specific documents
- Implemented evaluation metrics to assess faithfulness, relevance, and correctness
- Integrated the system into a workflow for summarizing and contextualizing technical analyses
- Containerized and operationalized the pipeline using Docker, with CI/CD automation via GitHub Actions and an interactive Streamlit interface
Methods & tools:
Python, LangChain, FAISS, Transformers, DeepEval, Docker, Github Actions
Outcome:
Improved accessibility and interpretation of complex modeling outputs for
technical and non-technical audiences.
🌆 NYC Climate Vulnerability Modeling
Focus: Spatiotemporal modeling · Deep learning · Simulation
Led the development of high-resolution spatiotemporal models to analyze population vulnerability to climate impacts across New York City.
Key contributions:
- Designed spatial feature engineering pipelines for demographic data
- Applied deep learning (LSTM-based models) to capture temporal and spatial dynamics
- Produced high-resolution projections by age and race to support climate vulnerability analysis in NYC
Methods & tools:
Python (rasterio, multiprocessing), R (sf, terra, tidyverse), TensorFlow
Outcome & impact:
Generated decision-ready insights for climate adaptation planning; work accepted
for publication (in-press) in Proceedings of the National Academy of Sciences (PNAS).
🔗 Project Link 🔗 GitHub Repository
🦠 AI-enabled COVID-19 forecasting
Focus: Predictive analytics · Time Series Analysis · Public Health
Co-led a collaborative effort to develop AI-enabled forecasting models for weekly COVID-19 incidence rates at the U.S. county level across multiple forecast horizons, with explicit uncertainty quantification..
Key contributions:
- Designed end-to-end machine learning workflows spanning data ingestion, feature engineering, model training, and validation
- Developed and evaluated several AI-based forecasting models using gradient boosting and deep learning
- Integrated spatial and temporal dimensions to support regional planning
- Implemented uncertainty-aware evaluation across short- and medium-term forecast horizons
Methods & tools:
XGBoost, LSTM models, scikit-learn, TensorFlow, time-series feature engineering,
uncertainty evaluation
Outcome & impact:
- Ranked among the top 5 most accurate forecasting models in the U.S. COVID-19 Forecast Hub
- Delivered improvements of up to 50 cases per county relative to the COVIDhub ensemble used by CDC decision-makers
- Contributed to real-time forecasting efforts supporting national public health planning
🧩 Common Themes Across Projects
- Emphasis on predictive and simulation-based modeling
- Integration of spatial and temporal context
- Explicit treatment of uncertainty and scenarios
- Focus on decision support, not just prediction accuracy